ai_security▌

Sweep log

The raw record of the crawler. Rejected candidates stay in, with the reason. That is the whole trust model.

# Sweep log

Raw record of every automated sweep, newest first. This file is rendered publicly at `/log/`.

Format:

```
## YYYY-MM-DD HH:MM UTC — sweep #N
+added:    <title> — <primary source URL>
~updated:  <title> — <what changed and why>
-rejected: <candidate title> — <reason, referencing the bar in sweep/SPEC.md>
```

The feed starts empty on purpose. No entries are backfilled after the fact — the first
real sweep is where this log begins.

Transparency note, 2026-11-09: sweeps #1–#2 were drafted by the sweep agent in Vietnamese and
translated into English afterwards, because this page is public and its readers are English
speakers (spec §6). Same lines, same URLs, same reasons — nothing added, removed or re-decided.
The agent's questions to the maintainer used to sit in this record; they now live in
`data/questions.md`, which is not published.

## 2026-11-08 18:40 UTC — sweep #1

Sources swept: OpenAI news, Anthropic news, Google Project Zero, Microsoft Security Blog,
Meta news, AWS Security, Unit 42, Embrace The Red, Daily Swig, BleepingComputer, arXiv
cs.CR+cs.CL (jailbreak/prompt-injection/red-team/model-extraction/agent-security), GitHub
advisories (llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch + GHSA DB + NVD
keyword LLM/AI), Hacker News (>=150 pts), HuggingFace blog, JFrog Research, Claude Code /
Codex / Cursor / Copilot changelogs. OpenAI pages 403 for datacenter curl → validated via
invisible-dots browser (spec §1). All 17 items below fetched with title/ID present in body.

+added:    WorkOS breach by researcher's ARTEX agent after sandbox escape — https://workos.com/security/bulletins/update-for-october-30th-2026-data-security-incident
+added:    OpenAI GPT-6 Cyber launch (81h old — exception: vendor's own first announcement + primary; BC corroboration 200) — https://openai.com/index/introducing-gpt-6-cyber/
+added:    OpenAI x Hacktron vulnerability-partnership — https://openai.com/index/partnering-with-hacktron/
+added:    Microsoft: agentic AI tooling used to find/abuse exposed secrets — https://www.microsoft.com/en-us/security/blog/2026-11-05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/
+added:    vLLM GHSA-mwgq-9582-w2vf response amplification — https://github.com/vllm-project/vllm/security/advisories/GHSA-mwgq-9582-w2vf
+added:    PwnEscape Part 2: Firecracker guest-to-host escape to host root (HN 285 pts) — https://pwn.ai/blog/firecracker-escape
+added:    CVE-2026-123653 DeepSeek-Reasonix doc-ingestion -> LLM prompting — https://github.com/advisories/GHSA-46pp-w572-hw78
+added:    CVE-2026-106020 Johnson Controls EasyIO LLM prompting handling — https://github.com/advisories/GHSA-g936-836g-r2pg
+added:    arXiv 2611.07640 indirect prompt injection / task-tampering benchmark (Nov 5, first-publication exception) — https://arxiv.org/abs/2611.07640
+added:    arXiv 2611.08281 Multimodal-SEA agent security benchmark (exception as above) — https://arxiv.org/abs/2611.08281
+added:    arXiv 2611.07370 cryptanalytic model extraction breaks TraintoDefend (exception as above) — https://arxiv.org/abs/2611.07370
+added:    arXiv 2611.07530 JailBound FOL-guided jailbreak evaluation (exception as above) — https://arxiv.org/abs/2611.07530
+added:    arXiv 2611.08226 'Blind Compliance' coercive control in jailbreaking (exception as above) — https://arxiv.org/abs/2611.08226
+added:    arXiv 2611.07586 CoT monitoring vs covert misalignment (exception as above) — https://arxiv.org/abs/2611.07586
+added:    CVE-2026-122160 AIlice untrusted model load, public PoC (Nov 5, first-reported + primary) — https://github.com/advisories/GHSA-6wf7-9xfh-x657
+added:    Cocos AI attestation gaps CVE-2026-119344/45 (NVD record; detail page is JS-rendered, ID verified via NVD API JSON; ACM paper + v0.9.0 release attached) — https://nvd.nist.gov/vuln/detail/CVE-2026-119345
+added:    CVE-2026-77604 code-graph-rag MCP read_file path traversal — https://github.com/vitali87/code-graph-rag/security/advisories/GHSA-vvr2-h2jp-838m

-rejected: "Anthropic IPO buyers see valuation as low as $1.1T" (HN 318) — funding news, not security (spec §3)
-rejected: "Cuban: Facebook knowingly runs AI deepfake scam ads, $16B fraud" (HN 515) — no primary source; opinion without PoC/repo/CVE (spec §3)
-rejected: "Alpine Linux advisory: OpenSSH password-auth bypass" (HN 627) — serious, but not AI security; out of niche
-rejected: "Fake security tools on GitHub push backdoors and infostealers" (BC) — generic malware supply chain, no AI factor
-rejected: "Fake OpenAI, SpaceX events used to steal accounts" (BC) — phishing that uses AI brands as bait, not an AI flaw
-rejected: "Every Mac ships an undocumented family-member classifier" (HN 814) — privacy research, not security
-rejected: "Attacks on Europe's refrigeration systems" (Unit 42 + FBI/CISA advisory) — ICS, no AI factor
-rejected: "Visual Studio Copilot remote code execution" (Embrace The Red, Oct 13) — >72h and not first-reported (circulating since Oct 13)
-rejected: "Jailbreaking my car's navigation system" (HN 952, Nov 2) — >72h, already reported widely
-rejected: vLLM GHSA batch Nov 2 (q9v9-qpjp-fc24, qrj2-j723-m6w8, 5r3c-5x4v-8gjh) — >72h; this log forbids backfilling
-rejected: "Making HTTP header injection critical with one prompt" (PortSwigger Research, pub 2026-11-04 15:12) — ~100h > 72h, published since Nov 4
-rejected: NVD keyword-AI noise: Dokan WooCommerce x3, HCL AION headers/open-redirect batch, Everest Forms, CTX Feed — generic web bugs, "AI" only in the product name
-rejected: CVE-2026-123336 YouTrack attachment-limit bypass via AI assistant — generic bypass, AI is only the route
-rejected: CVE-2026-48705 MONAI PydicomReader — an AI toolkit, but a generic DICOM reader bug

Swept, zero items inside the window: Project Zero (latest Oct 27), Meta news, AWS, HuggingFace
blog, Anthropic (latest post is the Opus 5.5 product announcement), Codex/Copilot/Cursor
changelogs (no security line within 72h), llama.cpp/transformers/langchain/crewai/pytorch (no
new advisory since Oct 27), JFrog Research (JS-rendered page, curl cannot list posts — asked as
Q3b in `data/questions.md` whether JFrog is worth sweeping through the browser).

## 2026-11-09 00:10 UTC — sweep #2

Sources swept: OpenAI/Anthropic/Meta/AWS/MSFT/Unit42/HF-blog/HF-malware-scanner feeds, Google
Project Zero, BleepingComputer (security feed), PortSwigger Research (Daily Swig feed dead — Q1
still open), JFrog Research, Techmeme, HN ≥150pts (72h), arXiv cs.CR+cs.CL listing (API export
returned 429 for the whole sweep → swept the /list/recent listing pages instead, counting
pagination), GitHub advisories (6 large AI repos + global DB desc-scan + pip/npm/go slices +
affects=transformers,langchain,vllm,torch,llama-cpp-python,ollama,mcp,openai,anthropic), NVD
keyword LLM/AI, Claude Code / Codex / Copilot / Cursor changelogs. Every URL was fetched in
this session.

+added:    Microsoft: Storm-3358 uses ARTEX against the financial sector (vendor primary, body 200) — https://www.microsoft.com/en-us/security/blog/2026-11-06/artex-abuse-expands-as-cybercriminals-target-financial-sector/
+added:    Claude Code v2.1.319 fixes sandbox escape when creating a worktree (release notes, no CVE) — https://github.com/anthropics/claude-code/releases/tag/v2.1.319
+added:    Apiiro: 'Don't Ask Again' writes commands containing passwords into a committed .claude/settings.local.json, 1,122 public files (HN 578; blog 403 for curl & JS-rendered — body not readable directly; corroborated via the authors' own HN announcement thread [spec §2] + Wayback snapshot 200; tldr contains only details seen in the fetched thread) — https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again
+added:    CVE-2026-18484 IBM ContextForge MCP Gateway tool-name collision DoS (GHSA 200, CVSS 6.5) — https://github.com/advisories/GHSA-2hgv-xvmf-mw3c
+added:    arXiv 2611.07222 SPILLWAY — separates target-admission from sanitization as prompt-injection defence for tool-using agents (pub 11-04, first-publication 11-06, the paper is the primary source) — https://arxiv.org/abs/2611.07222
+added:    arXiv 2611.08325 decision-model retrieval gate blocks injected instruction + pollution (pub 11-05) — https://arxiv.org/abs/2611.08325
+added:    arXiv 2611.06969 interactive backdoor via poisoning of fine-tune dialogue-LLM data (pub 11-04, first-publication 11-06) — https://arxiv.org/abs/2611.06969
+added:    arXiv 2611.06412 3 practical defences against backdoors from third-party data in an FR pipeline (announced 11-05, first-reported + primary; backdoor was not in the sweep #1 keyword list) — https://arxiv.org/abs/2611.06412

-rejected: "Goldman Sachs is pulling Claude Code" (HN 542) — source is a content-farm site, original is a paywalled Bloomberg story; a business decision, not a security finding (spec §3)
-rejected: "OpenAI raises $11B at $1.25T" (HN 179) + Anthropic IPO (HN 318, repeat) — funding, not security (spec §3)
-rejected: MaxxBench "Nice DeepSWE score. But is it true?" (HN 385) — about benchmark honesty, not security
-rejected: "Person of Interest" (krthr.co, HN 284) — opinion essay, no PoC/repo/CVE (spec §3)
-rejected: "Upload Your Ram" (HN 269) — a donated-RAM experiment, not a security finding
-rejected: Coldframe honeypot (HN 276) — a fridge honeypot, outside the AI niche
-rejected: Nvidia board investigates spy plot (HN 333) — corporate/espionage, no AI security angle
-rejected: "Every Mac family-member classifier" (HN 844) — already rejected in sweep #1 (privacy, not security)
-rejected: Codex v0.167.0 "MXC sandbox" line — general hardening, the release notes publish no security fix (spec §3: no evidence)
-rejected: GHSA IBM Turbonomic kubeturbo "agent" x3 — not an AI product, only the word "agent" matched
-rejected: arXiv 2611.07977 UAV interference detection — wireless, not AI security
-rejected: arXiv 2611.08155 AI-native semantic communications vulnerabilities — telecoms, outside the spec §2 arXiv keyword list
-rejected: arXiv 2611.05779 Private Vulnerability Testing for User Oversharing — privacy research, out of niche
-rejected: BC security items Nov 6-8 (SonicWall SSL-VPN MFA bypass, QNAP Plex2Shell, Chainalysis, Iris Ohyama, D-Link zero-days, MetaMask) — generic security, no AI factor
-rejected: Unit 42 "Europe's refrigeration" — already rejected in sweep #1 (ICS, no AI)
-rejected: latest PortSwigger Research 11-04 (already rejected in #1), latest JFrog 11-03 — outside the 72h window

Method notes (not policy questions):
- MCP invisible-dots DIED mid-session (dropped out of the registry) → no way around sources that
  block datacenter IPs, which spec §1 assumes exists. Apiiro was handled with corroboration
  (spec §3) instead of the browser.
- GitHub global advisory API: the `published=` filter is ignored and `page=` does not paginate
  (it returns page 1). What was swept: the 100 newest items (all of Nov 8) + pip/npm/go ecosystem
  slices + affects= for the large AI packages (0 hits in window). Remaining gap: global items
  published Nov 6–08 that belong to no slice and no AI repo — small risk of a miss, recorded here.
- llama.cpp/vLLM/transformers/LangChain/CrewAI/PyTorch: 0 new advisories in window (latest: vLLM
  11-02, already rejected in #1). Latest P0 10-27. Latest Anthropic news 10-12. HuggingFace: no
  security post in the feed. Copilot/Cursor changelog: 0 security-fix entries in window.

## 2026-11-09 12:00 UTC — sweep #3

Sources swept: OpenAI/Anthropic/Meta/AWS/MSFT/Unit42/HF-blog(incl. malware-scanner posts)/googleblog feeds,
Project Zero (homepage listing), BleepingComputer, PortSwigger Research, Embrace The Red, Daily Swig,
Techmeme, HN >=150pts (72h window, 62 stories), arXiv cs.CR+cs.CL /list/recent pages 1-2 (Mon 9 + Fri 6
announcement groups fully covered incl. the Fri-6 cs.CL page 2), GitHub advisories (6 large AI repos
reviewed+unreviewed, global DB created/updated/unreviewed slices, affects= AI packages), NVD keyword
LLM + AI (Nov 6 12:00 - Nov 9 12:00), ReversingLabs RSS, JFrog Research, Claude Code / Codex / Copilot
(CLI releases + github.blog changelog feed) / Cursor changelogs. Every URL below was fetched in this
session; arXiv titles verified present in each abstract page body. No datacenter-blocked source needed
the invisible-dots browser this round (OpenAI RSS and anthropic.com both answered curl).

+added:    Tracecat SAML signature-check bypass -> unauthenticated superadmin takeover (GHSA 200, body verified; NVD GET 200 corroboration; CVSS 6.6 High, MAJOR pending human review) — https://github.com/TracecatHQ/tracecat/security/advisories/GHSA-6g53-h6fh-c3pj
+added:    fast-agent serve default 0.0.0.0, no auth -> unauthenticated OS command execution with --shell (GHSA 200, body verified; CVSS 9.8) — https://github.com/evalstate/fast-agent/security/advisories/GHSA-9vhv-g5hf-m7m7
+added:    arXiv 2611.09734 GuardrailAgent adaptive multi-agent jailbreak defense (announced 11-09) — https://arxiv.org/abs/2611.09734
+added:    arXiv 2611.09721 MemoryDojo poisoning through LLM agent memory (announced 11-09) — https://arxiv.org/abs/2611.09721
+added:    arXiv 2611.09375 Cutpurse explanation-guided model extraction on tabular data (announced 11-09) — https://arxiv.org/abs/2611.09375
+added:    arXiv 2611.09348 reference-answer poisoning backdoors in agent skill optimization (announced 11-09) — https://arxiv.org/abs/2611.09348
+added:    arXiv 2611.09113 Authorization Drift Benchmark for multi-step tool-using agents (announced 11-09) — https://arxiv.org/abs/2611.09113
+added:    arXiv 2611.08845 Jev probabilistic safety guardrail empirical study (announced 11-09) — https://arxiv.org/abs/2611.08845
+added:    arXiv 2611.08745 runtime detection of malicious ML models beating static scanners (submitted 11-05, announced 11-09, first-publication exception as in sweep #2, the paper is the primary source) — https://arxiv.org/abs/2611.08745
+added:    arXiv 2611.08811 thought-injection attacks on reasoning agents (announced 11-09) — https://arxiv.org/abs/2611.08811
+added:    arXiv 2611.08752 AttentionSwitch jailbreak defense via sparse attention-head control (announced 11-09) — https://arxiv.org/abs/2611.08752
+added:    arXiv 2611.08711 RL-ATTACK black-box adversarial transformations against LLMs (announced 11-09) — https://arxiv.org/abs/2611.08711
+added:    arXiv 2611.08780 guardrail erosion in biomedical fine-tuned LLMs, paired audit (announced 11-09) — https://arxiv.org/abs/2611.08780
+added:    arXiv 2611.07611 LLM-planned honeypots routing autonomous attackers to a shadow copy (submitted 11-05, announced 11-06, first-publication exception as in sweep #2 — sweep #2's arXiv pass did not report it, log has no prior decision) — https://arxiv.org/abs/2611.07611

-rejected: "Goldman Sachs is pulling Claude Code" (HN 793) — repeat of sweep #2 rejection: content-farm rewrite of a paywalled Bloomberg story, business decision not a security finding
-rejected: "OpenAI raises $11B at a $1.25T valuation" (HN 188) — funding news, not security (spec section 3)
-rejected: Jensen Huang affair letter + Nvidia board statement (HN 723/471, Techmeme) — corporate news, no AI security content
-rejected: "Is Opus 5.5 nerfed? Thinking patterns changed on Nov 3" (HN 236, tweet) — opinion claim, no PoC/repo/CVE (spec section 3)
-rejected: "Unlocking any car from X brand in most of the world" (shubs.io, HN 416) — vehicle security, not AI security, out of niche
-rejected: Scamwich restaurant payment skimming (HN 170) — generic fraud, no AI factor
-rejected: Llama-modchip, 284B DeepSeek on 64GB RAM (HN 432) — inference/hardware hack, not a security finding
-rejected: "Every Mac family-member classifier" (HN 856), MaxxBench (HN 399), "Person of Interest" (HN 303) — re-posts of items already rejected in sweeps #1/#2 (privacy / benchmark honesty / opinion without evidence)
-rejected: Techmeme "prosecutors: businesses lost $129M+ to AI voice-clone scam calls" — law-enforcement roundup via secondary press, no primary source reachable (spec section 3)
-rejected: grok2api GHSA-8xq4-4x26-rxrw (CVE-2026-124574, Low, pub 11-09) — LLM gateway product but a generic missing-authorization bug in a video-job retry handler; AI is only the product name (YouTrack precedent)
-rejected: PraisonAI / PraisonAIAgents GHSAs — advisories published Apr 10; the Nov 8 change is a metadata update, not a first report (spec section 3)
-rejected: NVD keyword-AI noise Nov 6-8: HCL AION GenAI header/open-redirect batch (~24 CVEs), WordPress "AI" plugin XSS/privilege batches (Kubio, JS Help Desk, TranslatePress, Directorist, SEOPress, weDocs, AutomatorWP, GamiPress, Everest Forms), WWBN AVideo plugin/AI path traversal, itsourcecode ARS — generic web bugs, "AI" only in the product name (precedent sweeps #1-2)
-rejected: Apache Airflow Nov 9 batch (~18 CVEs) — keyword artifact ("AI" inside "Airflow"), workflow engine not AI security
-rejected: Copilot CLI v1.0.106-1.0.108 notes (sandbox hardening, terminal escape-sequence rendering) and Codex v0.169.0-alpha.2/3 — no advisory/CVE/security-fix line (Codex precedent sweep #2)
-rejected: AWS "Addressing protocol security gaps in multi-tenant Kubernetes" — k8s multi-tenancy isolation, no AI factor
-rejected: Unit 42 refrigeration post re-listed in feed (already rejected #1/#2, ICS no AI); BleepingComputer Nov 7-9 (Iris Ohyama breach, $133M French bank heist, gift-card scam ads) — generic security, no AI factor
-rejected: arXiv outside the spec keyword list: 2611.08985 dataset watermarking (provenance), 2611.09326 moral-decision benchmark, 2611.08651 fact-checking dataset, 2605.16336 copy-paste watermarks (copyright), 2611.08237 math benchmark, 2611.08170 agentic scambaiting (anti-scam), 2611.07403 vision-controller watermarking (CPS) — not AI security

Swept, zero items inside the window: OpenAI (latest 11-06: Hacktron already ingested; the wet-lab agent
eval post is capability research, not security), Project Zero (latest 10-27), Anthropic (newest security
post is distillation-defenses 10-12; verified via curl, no browser needed this round), Meta news,
HF blog/malware-scanner (latest is ToyBlocks, robot learning), Google security blog, ReversingLabs RSS,
Embrace The Red, PortSwigger, Cursor changelog (product features only), Claude Code releases (latest
v2.1.319 already ingested), llama.cpp/vLLM/transformers/LangChain/CrewAI/PyTorch per-repo GHSAs
(0 new in window), global GHSA reviewed list updated >= 11-06 12:00 (34 entries, only the PraisonAI
metadata update, see rejections). Daily Swig feed still dead (Q1 open), JFrog still JS-rendered (Q3b open).

Method notes (not policy questions):
- nvd.nist.gov answers HEAD with 403 but GET with 200 — validator step 1 must use GET, not curl -I.
- GitHub /advisories still ignores page= (page 2 identical to page 1), and repo-level *unreviewed*
  GHSAs (this sweep's Tracecat + fast-agent, like vLLM mwgq in #2) 404 in the API and miss every list
  endpoint while their HTML pages are live 200. AI coverage therefore = per-repo endpoints (reviewed
  + unreviewed) + affects= slices + NVD keyword LLM/AI, which caught both.
- arXiv /list/recent truncates a date group across pages (cs.CL Fri-6 showed 22 of 94 on page 1) —
  this sweep fetched skip=100 too, so both categories are complete for the window.

## 2026-11-10 00:40 UTC — sweep #4

Sources swept: OpenAI (news RSS; openai.com HTML answered 403 to datacenter curl this round),
Anthropic news, AWS Security blog, Microsoft Security blog, Google Project Zero, googleblog,
Meta news, Unit 42, HuggingFace blog, PortSwigger Research (= the live successor of the dead
Daily Swig feed, Q1 open), Embrace The Red, ReversingLabs blog RSS, JFrog blog (still
JS-rendered, Q3b open), BleepingComputer (security feed + AI section), Techmeme (feed.xml),
github.blog changelog feed, HN stories >=150 pts (window Nov 9 12:00 - Nov 10 00:01 UTC),
arXiv cs.CR+cs.CL /list/recent pages 1-2 incl. skip=100, GitHub advisories
(llama.cpp/vLLM/transformers/LangChain/CrewAI/PyTorch/Tracecat/fast-agent per-repo endpoints +
global 100-newest + affects= AI packages), NVD keyword LLM / "prompt injection" / AI
(published Nov 9 12:00 - Nov 10 00:30, 0 results), Claude Code / Codex / Copilot CLI release
feeds. Every URL below was fetched in this session.

+added:    Fireworks AI bulletin update: the October incident also hit secrets, datasets and incomplete model weights for a limited set of accounts; LLMJacking attribution + second intrusion path via a third-party cloud provider (bulletin 200, body verified; Unite.AI corroboration 200) — https://fireworks.ai/security/bulletins/october-2026-security-incident-update
+added:    CVE-2026-88968 Claude Code Edit-tool TOCTOU: sandboxed process swaps a validated directory for a symlink, unsandboxed Edit overwrites files outside the workspace (GHSA 200, body verified, patched; reported by Sunder Singh via HackerOne) — https://github.com/advisories/GHSA-5xjx-9qx6-jcrv
+added:    CVE-2026-102895 GPUStack loopback auth bypass: same-host reverse-proxy shape or forged X-GPUStack-Real-IP gives unauthenticated network admin over LLM serving clusters (GHSA critical, fixed v2.2.0) — https://github.com/advisories/GHSA-gj9m-f8ph-pmhr
+added:    Kubeflow Pipelines Nov 9 batch: empty-namespace upload endpoints skip authorization -> unauthenticated injection into shared pipelines (image/command swap other tenants execute); same batch cross-tenant cache output substitution, global-pipeline RBAC bypass, unvalidated pipelineURL fetch; no patched versions listed (4 GHSA pages 200, bodies verified) — https://github.com/advisories/GHSA-rjr3-43j3-4q8g
+added:    CVE-2026-87008 code-graph-rag structural MCP tools follow symlinks out of the project root (lexical containment check); second advisory in the repo after CVE-2026-77604 already ingested — https://github.com/advisories/GHSA-85gg-2gfq-q95m
+added:    Claude Code v2.1.321 release notes: MCP OAuth secrets sent to another host on redirect, read-only Bash outside the sandbox, wrong-folder write permission checks, cp/mv deny-rule gaps (release notes, no CVE; same pattern as the sweep #2 v2.1.319 item) — https://github.com/anthropics/claude-code/releases/tag/v2.1.321
+added:    Codex app Remote SSH launches the remote app-server on the user's forwarded SSH agent socket (PR by abhinav-oai + maintainer comment confirming the app already does this; HN 205 pts argues remote agents can then use local SSH credentials) — https://github.com/openai/codex/pull/28131
+added:    AWS: multi-layer write validation against memory poisoning in Bedrock AgentCore (IAM CreateEvent gating, namespace validation, consolidation prompts, ApplyGuardrail + embedding semantic-drift signal) — https://aws.amazon.com/blogs/security/help-defend-against-memory-poisoning-in-amazon-bedrock-agentcore-with-multi-layer-write-validation/
+added:    Microsoft FORGE Lab: Hyenas agentic scanner (frontier MDASH) finds memory-safety issues incl. safe-Rust undefined behavior in std/core; three misplaced-trust patterns — https://www.microsoft.com/en-us/security/blog/2026-11-09/forge-lab-uses-agentic-tools-to-identify-memory-safety-issues-in-rust/

-rejected: OpenAI "Control Room" (Nov 9) — internal research-workload tracking/stop system per the vendor's own summary: safety operations, not an attack/defense or vulnerability finding (precedent: wet-lab agent eval post, sweep #3); the page 403s to datacenter curl and there is no second source to corroborate a body nobody read (spec section 3)
-rejected: "Google Gemini routes personal Drive files and emails to human annotators" (HN 308, repo with testing) — privacy/policy change with no attacker and no integrity flaw (precedent: every-Mac classifier, sweep #1)
-rejected: Anthropic Fable 5.2 / Mythos 5.2 launch + pricing + benchmark posts (HN 853/254, Techmeme), Meta Muse Ember (HN 225), "Claude Fable 5.2 in GitHub Copilot" changelog — product releases, no security content
-rejected: Copilot CLI v1.0.108/v1.0.109-0 notes (MCP server restart after /cd to follow denied paths, symlink confirmation on /share, hidden-path review) — permission/sandbox hardening lines with no advisory or CVE (precedent sweeps #2/#3)
-rejected: Codex 0.169.0-alpha.6/7 + 0.167.0-alpha.4.1 — empty release notes (Codex precedent sweep #2)
-rejected: tsParticles MCP-server trio CVE-2026-76892/76893/63655 (session monopoly, malformed-JSON crash, body-buffering DoS) — generic HTTP DoS on an animation library, "MCP" is only the surface (generic-bug precedent)
-rejected: fastapi-code-generator x3 critical CVE-2026-63469/63470/63471 — codegen template injection from OpenAPI specs, no AI factor
-rejected: ezBookkeeping CVE-2026-86041 (X-Forwarded-For bypass of the MCP/API-token allowlist) — generic header-trust bug, MCP only as a feature name
-rejected: zizmor CVE-2026-102289 debug-log credential echo — generic logging bug in a CI scanner, no AI factor; same class for the rest of the Nov 9 global batch (Telepresence, Pocket ID, traQ, isomorphic-git, FileGator, OmniAuth, Gradle Shadow, JSON-java, ag-psd, zt-zip, DDEV, croc, Moleculer, Velociraptor, Kiwi TCMS, Urwid, esm.sh, Remark42)
-rejected: Kubeflow Pipelines CVE-2026-88985 (SQLi via pagination token) + CVE-2026-88986 (reflected XSS in error responses) — generic web bugs, kept out of the accepted multi-tenant item
-rejected: fast-agent GHSA-9vhv-g5hf-m7m7 and code-graph-rag GHSA-vvr2-h2jp-838m reappearing as "published" Nov 9 ~22:00 UTC — unreviewed-to-reviewed promotion in the global DB of items already ingested in sweeps #3/#1, no new content
-rejected: OpenAI "ChatGPT Enterprise in AWS Marketplace" + CodeRabbit + Kuken customer stories — business/marketing posts
-rejected: BleepingComputer Nov 9 (Vanderbilt BYOD ransomware, SEGA newsletter breach, N-able zero-days, York University, France/Iran supermarket attack, Beget wiper, Binance/ShinyHunters, SSHole exploitation, FBI agent crypto theft, MEXC sanction ban) — generic security, no AI factor
-rejected: Techmeme Nov 9: Apple "intelligence usage packs" (product), Jensen Huang affair + Nvidia board items (corporate), Binance memecoin laundering (crypto forensics), Megaspeed chip-smuggling transfer (export enforcement), French supermarket refrigeration (ICS, no AI, precedent sweeps #1-3)
-rejected: HN repeats of earlier decisions: Goldman Sachs pulls Claude Code (882), Is Opus 5.5 nerfed (717), shubs car unlock (597), Upload Your RAM, MaxxBench, "Person of Interest", Coldframe, llama-modchip, AI-bubble essay; Apple grey-list spelling (549) is new here — product/policy behavior, not security
-rejected: "Agent memory needs its OpenTelemetry moment" (HN 189) — standardization essay, opinion without PoC/repo/CVE (spec section 3)
-rejected: HN non-security window items: USPS ZIP polygons, E-means, satellite photos, Pi5 pricing, RTR payments, H100 pricing, Ethereum lawsuit, "I saw your HN comment", magenta solar panels, Wikipedia essay, paper airplanes, malloc, Harvest post, DMARC policy, Outlook, Tesla, Dux, KetchUp, Chimera, OpenStick, Turbo Pascal, Pratchett, CPU dies, Electron/Discord, RAM-email prank, Starship 15, MS Research AI Frontiers shutdown (org news) — outside AI security

Swept, zero items inside the window: llama.cpp/vLLM/transformers/LangChain/CrewAI/PyTorch/
Tracecat per-repo advisory endpoints (0 new; latest vLLM still Nov 2, rejected in #1), Project
Zero (latest Oct 27), Meta news (Nov 9 post is teen-account privacy), googleblog (latest Nov 3),
Unit 42 (latest Nov 6, already rejected), HuggingFace blog (Nov 9 posts are dataset/pruning
research), PortSwigger Research (latest Nov 4, rejected in #1/#2), Daily Swig (feed still dead,
Q1 open), Embrace The Red (latest post still the Oct 13 Copilot RCE rejected in #1),
ReversingLabs (latest Nov 4), JFrog (JS-rendered, Q3b open), github.blog changelog (Nov 9:
Fable 5.2 on Copilot product post + secret-scanning workflow-log feature, 09:01 UTC i.e. also
before this window's cutoff), NVD keyword LLM/prompt-injection/AI (0 published in window),
arXiv: the latest announcement group is still Mon Nov 9 + Fri Nov 6, both fully covered by
sweep #3; the Tue Nov 10 group publishes at ~01:00 UTC today, i.e. it is sweep #5's work.

Method notes (not policy questions):
- The global GHSA DB "published" field reflected a Nov 9 ~21:00-23:30 UTC unreviewed->reviewed
  promotion batch, not first disclosure (it re-stamped already-ingested vvr2 and fast-agent).
  Triage therefore dedups against data/items.json, and the accepted GHSA items above are new
  content, verified by reading each advisory body.
- NVD detail pages now answer 200 but the CVE id is not in the served HTML (JS-rendered), so
  NVD pages are no longer usable as corroboration; corroboration here is same-topic GHSA
  cross-links and vendor press (Unite.AI 200 with Fireworks detail).
- invisible-dots was not started this round: the only datacenter-blocked source was openai.com
  HTML, and its entire Nov 9 feed was rejected on scope from the vendor's own RSS summaries,
  so no candidate needed a body check through the browser.
## 2026-11-10 12:00 UTC — sweep #5

+added:    arXiv MetaPI: indirect prompt injection benchmarks derived from agent utility benchmarks — https://arxiv.org/abs/2611.10026
+added:    arXiv: Mythos 5 attempted a real-person supply-chain attack during UK AISI cyber evaluations (alignment investigation report) — https://arxiv.org/abs/2611.10030
+added:    arXiv: limits of information-flow control enforcement for agentic systems (label creep) — https://arxiv.org/abs/2611.10395
+added:    arXiv: experience backdoors poisoning self-evolving LLM agents via distilled skills — https://arxiv.org/abs/2611.10542
+added:    arXiv: constructing and analyzing an MCP security vulnerability dataset — https://arxiv.org/abs/2611.11104
+added:    arXiv: coding agents propagate vulnerabilities in code bases over time — https://arxiv.org/abs/2611.11455
+added:    arXiv: prompt injection security mechanised relative to a reference task, with defences and verdicts — https://arxiv.org/abs/2611.11506
+added:    arXiv: scaling adaptive prompt injection attacks with agent swarms — https://arxiv.org/abs/2611.11861
+added:    arXiv: StealthyPoison — procedural poisoning of third-party agent skills in public registries — https://arxiv.org/abs/2611.12155
+added:    arXiv: shibboleth traps — fingerprinting the LLMs and harnesses behind agentic attacks — https://arxiv.org/abs/2611.12386
+added:    DeepMind AlphaProof Code formally verified Signal's core cryptographic protocol in four weeks — https://deepmind.google/blog/introducing-alphaproof-code/
+added:    Shimmy AI-inference server served its Inference API unauthenticated through 2.6.4 (CVE-2026-125087, GHSA-h7qf-5c8w-fgjc) — https://github.com/advisories/GHSA-h7qf-5c8w-fgjc
+added:    Hugging Face finetrainers model-index code injection with public exploit (CVE-2026-124588, GHSA-jx9r-269r-fp82) — https://github.com/advisories/GHSA-jx9r-269r-fp82
+added:    langchain4j-aideepin IDOR leaks other users' knowledge-base chunks, agent memory and knowledge graphs (CVE-2026-125386, GHSA-58f4-85vm-hv79) — https://github.com/advisories/GHSA-58f4-85vm-hv79
+added:    Google Gemini routes connected personal Drive files and Gmail mail to human annotators (repo research, 632 HN points) — https://github.com/thisisme6598/gemini_files

-rejected: LMDeploy SSRF URL-parser-disagreement advisory (GHSA-39wr-7q6h-cf68, CVE-2026-123818) — the GHSA itself has been public since 2026-09-16; only the NVD CVE entry is new (Nov 9 22:18). Outside 72h and not a first reporting (spec section 3)
-rejected: GHSA-wchh-mrv2-5h4q ezBookkeeping source-IP allowlist bypass via spoofed X-Forwarded-For on MCP/API-token endpoints — bookkeeping app, not an AI repo; the bug is a generic IP-allowlist spoof
-rejected: GHSA-qfxj-2x5m-gvvg "MCP Server for WordPress" plugin privilege escalation — a WordPress plugin web vuln found by Patchstack; MCP is the connectivity feature, not the affected AI stack (spec section 2 scope)
-rejected: GHSA-h7c4-q426-7jh9 / GHSA-4p8p-pjpq-2w7h / GHSA-447x-qqwf-m7xq tsParticles @tsparticles/mcp-server DoS trio — animation-library MCP server, DoS-only; page header reads "Published Jul 10, 2026 - Updated Nov 9, 2026", i.e. the Nov 9 promotion re-stamp from sweep #4's method note, not new disclosure
-rejected: GHSA-f58r-c62m-574f Apache Axis2/Java metadata disclosure (its "MCP catalog endpoints" wording matched the keyword) — Java web-services stack, not AI security
-rejected: GHSA-g498-8632-6hwv Apache Storm DefaultStateSerializer — keyword noise, not AI security
-rejected: Nov 9-12:00 GHSA batch keyword hits that are not AI: Hitachi Dynamic Link Manager fleet ("AI" inside "AIX"/"Link Manager"), Everest Forms "with AI" WordPress plugin, GV-Face/GV-Cloud/GV-Access hardcoded camera-system API keys, LimeSurvey XSS, plus the SAP NetWeaver / ASUS router / WordPress-plugin / D-Link publication batches — outside scope (spec section 2)
-rejected: OpenAI "Closing the AI guidance gap in education" (Nov 10 09:00) — education/product post, no security content; OpenAI Nov 9 feed already dispositioned in sweep #4 (incl. Control Room)
-rejected: Anthropic "Introducing Claude Fable 5.2 and Claude Mythos 5.2" (HN 660/978, posted Nov 9 18:01, still undispositioned when sweep #4 ran) — model release post, product launch not security; the security-relevant artifact of the same model line this sweep is the arXiv 2611.10030 incident report, accepted above
-rejected: HN Nov 10 window non-security: Strands Agents "Amazon saves millions replacing Sonnet" (370, business/engineering), CoreWeave bond yields (173, finance), "Locating the source of a mysterious beep" (217), satellite photo browser (375), Muse Ember model release (310, product) — outside AI security
-rejected: HN "Horible public key tricks" (182) and "Your TV can see through walls" (149) — security-adjacent but not AI; the second also below the 150-point bar
-rejected: HN "AI usage generates a fragile codebase" (66) — opinion, under the points bar, no PoC/repo/CVE (spec section 3)
-rejected: HN repeats of earlier decisions resurfacing on the front page: USPS ZIP polygons (637), Payments Canada RTR (538), E-means (419), magenta solar panels (323), Pi5 pricing, H100 pricing, Wikipedia essay, "I saw your HN comment" — outside scope
-rejected: Techmeme Nov 10: Anthropic IPO roadshow + $64.4B prospectus + "Jagged AGI" review (business/corporate), Jensen Huang affair and Nvidia board items (corporate), Meituan LongCat-2.5 open-weights release (product), SoftBank loss and GeniosAI raise (finance), BT/Starlink, China trade surplus — no security angle
-rejected: BleepingComputer Nov 10: Barracuda ESG root RCE in exploitation, Turla "Ceral" malware vs government Exchange, Chrome eighth exploited zero-day — generic security, no AI factor
-rejected: BleepingComputer "Fireworks AI discloses new breach after cloud provider hack" (Nov 10 02:49) + Notion "security incident involving a third-party AI vendor" statement (HN 67) — coverage of the already-ingested 2026-11-09 Fireworks incident (Notion confirmed as an affected customer; corroboration only, no new item)
-rejected: arXiv Nov 10 group outside scope: O-RAN/smart-grid/IoT IDS ML, wireless-PHY and physical-layer authentication, CAPTCHA-GAN breaking, cross-chain bridge post-mortems, code watermarking, DER grid commands, homomorphic UAV comms, Android manifest malware clustering — not AI-product security; plus replacement/cross announcements of pre-November papers (2502.x-2610.x ids) which are not first announcements
-rejected: arXiv Nov 10 candidates held out for signal density (same bar as sweep #4): CyberBinGym (agents turning patches into exploits), "Detecting Shadow AI" enterprise endpoint detector, SNAPRAG RAG knowledge-extraction attack, "The Devil Wears Markdown" skill-intent detection, Double Blind Evals, Chatter Attack LLM resource exhaustion, Honeyval, "When Coding Agents Get Arms" — real but secondary to the ten accepted
-rejected: Coding-agent security changelogs: Claude Code v2.1.322 (Vertex fast-mode only, no security note), Codex rust-v0.169 alpha releases (empty bodies), Copilot changelog (nothing new since the Nov 9 items dispositioned in sweep #4), Cursor changelog (latest Nov 6: voice chat, build live apps — product, outside window) — no security entries this window

Swept, zero new items in window: per-repo advisory endpoints llama.cpp / vLLM / transformers / LangChain /
CrewAI / PyTorch (latest vLLM still Nov 2, dispositioned in #1; all others unchanged), Google Project Zero
(latest Oct 27), Microsoft Security Blog (nothing since the Nov 9 FORGE post, ingested), AWS Security Blog
(nothing since Nov 9 19:05 AgentCore post, ingested), Meta news (Nov 9 teen-account post, rejected in #4),
Unit 42 (latest Nov 6 refrigeration report, rejected in #4), Hugging Face blog (Nov 9 dataset/pruning posts,
rejected in #4), Portswigger Research (latest Nov 4, rejected in #1/#2), Embrace The Red (latest Oct 13),
ReversingLabs blog (AI-coding-agents-screenshots post is Oct 13 per sitemap; latest post Oct 27), Signal
blog (latest Sep 28, checked after the AlphaProof lead), GitHub changelog (Nov 9 items dispositioned in #4),
Daily Swig (feed still dead: curl exit with no HTTP response, Q1 open), JFrog (still JS-rendered, Q3b open).

Method notes (not policy questions):
- The GitHub global advisories endpoint caps at 100 per page and ignored page=2 in this range (returned
  page 1 again), so the window was bisected into 11 sub-windows (three needed a second split; one 30-minute
  window still held 100 entries from the Nov 9 ~21:00-23:30 promotion batch). Union: 503 unique advisories,
  deduped against data/items.json. The tsParticles "Published Jul 10" header inside a published=2026-11-09
  result range independently confirms sweep #4's re-stamp finding.
- arXiv: the Tue Nov 10 group (predicted by sweep #4) was today's payload: 500 RSS entries across cs.CR+cs.CL,
  10 accepted. Each accepted abs page was fetched this session and the title verified in the page; "released"
  is the v1 submission date shown on the abs page (Nov 5-9), announced-for-first-time today per spec section 3.
- NVD keyword search Nov 9 12:00 -> Nov 10 12:30: "prompt injection" 0, "AI agent" 0, "large language model"
  1 (LMDeploy CVE-2026-123818, rejected on age, see above). NVD detail pages remain JS-rendered and were not
  used as source or corroboration.
- invisible-dots was not started: every source/corroboration URL returned 200 to direct fetch; no source
  blocked the datacenter IP this round. openai.com HTML still blocks bots; its RSS was used instead.

## 2026-11-11 00:00 UTC — sweep #6

Window: Nov 10 12:00 UTC -> Nov 11 02:40 UTC. Sources swept: OpenAI news (RSS), Anthropic news
(sitemap + page), Google Project Zero (feed moved to projectzero.google/feed.xml), Microsoft
Security Blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red, Daily Swig,
BleepingComputer, arXiv cs.CR+cs.CL RSS (re-fetched at 02:33 UTC), GitHub global advisories API
(window bisected, see method notes) + per-repo advisory endpoints (llama.cpp, vLLM, transformers,
LangChain, CrewAI, PyTorch), NVD keyword searches (prompt injection / AI agent / large language
model / LLM), Hacker News >=150 pts (Algolia API), Techmeme (feed.xml), Hugging Face blog,
ReversingLabs, JFrog, GitHub blog + changelog, coding-agent changelogs (Claude Code, Codex,
Copilot, Cursor). All 8 added URLs fetched this session, HTTP 200, title/ID present in body.

+added:    Railway full incident report: attacker specifically targeted customer AI-provider credentials for resale into distillation pipelines; env vars decrypted for at most ~100,000 workspaces; ~$740,000 unauthorized AI spend — https://blog.railway.com/p/incident-report-october-security-incident
+added:    PyTorch GHSA-v32g-mr63-3j83: torch.load(weights_only=True) arbitrary text-file read via auto-trusted traceback.FrameSummary (Torch 2.14, high, no patch listed) — https://github.com/pytorch/pytorch/security/advisories/GHSA-v32g-mr63-3j83
+added:    AG2 CVE-2026-113663 GHSA-2h4w-29cv-2cmv: newline injection in SandboxShellTool bypasses allowed=[...] and readonly=True, arbitrary shell exec via sh -c (high; ag2 fixed 1.1.1) — https://github.com/advisories/GHSA-2h4w-29cv-2cmv
+added:    Kong AI Gateway CVE-2026-92544 GHSA-jvch-w6fx-vccj: SQL injection in LLM usage counters via crafted model name (high) — https://github.com/advisories/GHSA-jvch-w6fx-vccj
+added:    Cocos AI (ultravioletrs/cocos) CVE-2026-119345 GHSA-ghwv-vrp2-2975: attested TLS keeps trusting a confidential AI workload after post-attestation state swap, TOCTOU (high; fixed 0.9.0) — https://github.com/advisories/GHSA-ghwv-vrp2-2975
+added:    vLLM GHSA-v97p-wpxp-px84: cross-user token data leakage with async scheduling + pipeline parallelism, v0.16.0-v0.20.2, fixed 0.21.0 (published Nov 9 17:29, never dispositioned by sweeps #1-#5, within 72h of GHSA publication) — https://github.com/vllm-project/vllm/security/advisories/GHSA-v97p-wpxp-px84
+added:    vLLM GHSA-fpch-3xrf-7q96: thinking-budget sparse state aliases onto a co-batched request during mixed SWAP, forcing reasoning-end tokens cross-session (CWE-488), v0.21.0-v0.27.1, fixed 0.28.0 — https://github.com/vllm-project/vllm/security/advisories/GHSA-fpch-3xrf-7q96
+added:    mcp-atlassian CVE-2026-124503 GHSA-8fch-hp3f-8mp8: path traversal in validate_safe_path, public exploit, fixed 0.24.1 — https://github.com/advisories/GHSA-8fch-hp3f-8mp8

-rejected: LiveDocs CVE-2026-63651 / GHSA-xf4p-x7v8-3x4p (SSRF in fetch_docs/introspect, prompt-injection chain; matched NVD keyword "AI agent" Nov 10 17:17) — the repo advisory itself is published 2026-07-01 (page shows that date); only the NVD CVE entry is new. Outside 72h and not a first announcement (spec section 3, same rule as the sweep #5 LMDeploy rejection)
-rejected: pgEdge RAG Server CVE-2026-125326 / GHSA-68g8-gh8f-7rhg (provider credential fragment disclosure via API errors) — advisory published 2026-09-03; only the NVD entry (Nov 10 18:17) is new. Same age rule as above
-rejected: pgEdge Go LLM Library CVE-2026-125328 / GHSA-8835-xmqj-354h (unauthenticated /v1/health pings providers with operator credentials) — advisory published 2026-09-03; only the NVD entry (Nov 10 19:21) is new. Same age rule
-rejected: TaxHacker CVE-2026-125749 (NVD keyword "LLM" hit) — IDOR in tax-app settings server actions; the LLM mention is a product feature, the bug is a generic cross-user authz flaw in a web app (spec section 2 scope; ezBookkeeping precedent from sweep #5)
-rejected: Dagu GHSA-5m6c-38r8-5w76 / CVE-2026-102943 — MCP is a bolt-on surface of a DAG/cron scheduler, not an AI product; read-only .yaml containment bypass with a viewer-role key (same treatment as the sweep #5 tsParticles and WordPress-MCP rejections)
-rejected: React GHSA-xqc2-xf4c-x3h3 unauthenticated RCE in React Server Components — severe, but a web-framework bug with no AI factor; outside scope (spec section 2)
-rejected: Firefox 160 GHSA-2f94-x79r-94p2 "Sandbox escape in the Agent component" — keyword noise; Mozilla's internal component name, not an AI agent
-rejected: Microsoft November 2026 Patch Tuesday import batch (GHSA promotion ~21:32-21:34 UTC plus the 18:31 and 15:33 import batches; ~1,235 flaws across Windows/.NET/Office/SQL Server/Android/Adobe/MongoDB/QNX/Thunderbird/AthonetOS; keyword hits were substrings of "maintain", "aggregate", "analysis") — generic platform security, no AI factor
-rejected: Mendix Studio Pro 11 pair CVE-2026-112136/CVE-2026-112137 — a real finding (the low-code IDE's bundled MCP Server binds its HTTP listener beyond loopback), but MCP is an accessory of a non-AI dev tool, same treatment as Dagu above; ManageEngine DDI Central CVE-2026-95651 — keyword noise ("Linux Node Agent" component, an unauthenticated GSS-TSIG keytab flaw in a DNS/DHCP/IPAM product)
-rejected: compose-lint workflow-command-injection trio (GHSA-8v59-qmwv-4p6q, GHSA-xcf7-88g6-7xr8, GHSA-hr4h-4c45-v4q4) and Rails Active Storage pair — CI linter and web-framework bugs; keyword substring noise
-rejected: GitHub blog "Git security vulnerabilities announced" (Nov 10 18:00) — Git client vulnerabilities, broad but no AI factor; outside scope
-rejected: AWS Security Blog "Reduce alert fatigue by customizing threat detection features in Amazon GuardDuty" (Nov 10 23:04) — ops how-to for triage; Bedrock/Cross-Region-Inference/AWS MCP Server appear as the feature implementation, no new disclosure or research (spec section 3 product-post policy)
-rejected: AWS Security Blog ProcessUnity GRC assessment refresh + "Digital sovereignty at re:Invent" session list — compliance/marketing, no security content
-rejected: OpenAI Nov 10 posts: Introducing Businesses in ChatGPT (10:00), Helping caregivers use AI in everyday life (12:00), Work at the Frontier freelance study (07:00) — product and society posts, no security content
-rejected: Anthropic "Anthropic launches roadshow for proposed initial public offering" (Nov 10) — corporate/business (sitemap lastmod; same policy as the sweep #5 IPO rejections)
-rejected: BleepingComputer window non-AI: Nitori 37.6M customer records, Patch Tuesday 1,235 flaws + Outlook zero-day, Xen Orchestra auth bypass exploited, Binance iOS wallet-secrets flaw, "hackers exploit trusted AWS access" cloud drift campaign, Windows 10 ESU KB5131096 — generic security, no AI factor (sweep #5 precedent)
-rejected: Techmeme window: Daylight $80M Series B for AI managed security (funding), Cubist/Point72 filing alleging an ex-analyst used Claude and Codex to take source code (legal dispute over insider IP, not an AI product security disclosure), Graphon agent-memory launch (product), Zen 6 deep dive, CoreWeave/OpenAI financials, Starlink Mini 2 — no AI-security angle
-rejected: HN Nov 10 12:00 -> Nov 11 00:30 >=150 pts (15 stories): all non-security or already dispositioned — AlphaProof Code (564, ingested in #5), Gemini 4 Carbon and Claude Whiteboards/Sheets/Data (product launches), Scion kei truck, Starlink Mini 2, Synaptics/DisplayLink, Val Town winding down, Zen 6, .name grandfathered, Blender 5.3, Ableton 13, .NET 11, smart-home-offline essay, Valencia visualizer
-rejected: arXiv 2611.03555 "MCP Tool Poisoning in Agentic AI Frameworks" (today's only keyword hit not dispositioned by sweep #5) — abs page shows v1 Nov 2, v2 Nov 9: a replacement announcement, not a first announcement (sweep #5 rule)
-rejected: Coding-agent security changelogs: Claude Code v2.1.322 (Nov 10 02:29, no security lines in notes), Codex rust-v0.167.1 (Nov 10 21:32, body is a compare link only) and 0.169.0-alpha.7/alpha.9 (empty bodies), GitHub changelog (Gemini 4 Carbon in Copilot, GitHub App support for Copilot seat management, C++ LSP auto-config, Dependabot alerts pausing on inactive repos, dashboard — product changes, no security fixes), Cursor changelog (still Nov 6) — no security entries this window

Swept, zero new items in window: Microsoft Security Blog (latest Nov 9 FORGE post, ingested in #5),
Google Project Zero (latest Oct 27), Unit 42 (latest Nov 6, rejected in #4), Embrace The Red
(latest Oct 13), Meta newsroom (Nov 9 teen-account post, rejected in #4), Hugging Face blog
(Nov 10: Hub-on-Google-Marketplace and Decision Models — product posts), OpenAI news RSS (nothing
security-related), per-repo advisory endpoints llama.cpp / transformers / LangChain / CrewAI
(nothing newer than June), Daily Swig (feed still dead: /daily-swig/news/rss returns an HTML app
shell with zero <item> entries — Q1 open), JFrog blog index (zero server-side article links —
Q3b open), ReversingLabs (rss.xml serves the Next.js HTML shell to curl; no newer post reachable
than the Oct 27 one noted in sweep #5).

Method notes (not policy questions):
- GitHub global advisories endpoint again caps at 100 rows/page; the window was bisected into six
  sub-windows against the import batches (15:33, 15:56-57, 18:31, 21:32, 21:33-34 UTC), with the
  12:00-15:32:59 slice returning 25 rows and fully enumerated. All keyword hits reviewed against
  the batch contents above.
- The per-repo vLLM endpoint listing now shows two advisories published Nov 9 17:29/17:30 UTC that
  sweep #5's per-repo check did not see ("latest vLLM still Nov 2"). Greps confirm neither GHSA id
  appears in data/items.json or in any earlier -rejected line, so both entered the feed on first
  ingestion, within 72h of the advisory's own publication. Repo-advisory detail pages 404 on the
  unauthenticated API, so each was validated via its HTML page (200, title in body).
- Railway vs Fireworks: sweep #5 ingested the Fireworks-side bulletin (2026-11-09 item). Today's
  item is Railway's own provider-side report — a separate vendor disclosure of the platform-side
  incident, added as its own item rather than updating the Fireworks entry.
- NVD keyword searches Nov 10 12:00 -> Nov 11 02:30: "prompt injection" 0, "large language model"
  0, "AI agent" 1 (LiveDocs, rejected on age), "LLM" 4 (Kong accepted; pgEdge x2 and TaxHacker
  rejected above). NVD detail pages remain JS-rendered and were not used as source.
- arXiv: the cs.CR/cs.CL RSS was byte-identical when re-fetched at 02:33 UTC — still the Tuesday
  group dispositioned by sweep #5; the Wednesday Nov 11 announcement (due ~01:00 UTC but not yet
  propagated to RSS) is the next sweep's payload, same hand-off pattern as #4 -> #5.
- invisible-dots was not started: every source and corroboration URL returned 200 to direct fetch
  (openai.com via its RSS; Techmeme via feed.xml, its homepage serving a JS shell to curl).

## 2026-11-11 12:00 UTC — sweep #7

Window: Nov 11 02:40 UTC -> Nov 11 12:00 UTC (picks up where sweep #6 stopped; includes the
Wednesday Nov 11 arXiv announcement sweep #6 handed off). Sources swept: OpenAI news (RSS),
Anthropic (sitemap), Google Project Zero (feed), Microsoft Security Blog, Meta newsroom,
AWS Security Blog, Unit 42, Embrace The Red (feed path corrected, see method notes), Daily Swig,
BleepingComputer, arXiv cs.CR+cs.CL RSS, GitHub global advisories API (window fully enumerated,
see method notes) + per-repo advisory endpoints (llama.cpp, vLLM, transformers, LangChain,
CrewAI, PyTorch), NVD keyword searches, Hacker News >=150 pts (Algolia API), Techmeme (feed.xml),
Hugging Face blog, ReversingLabs, JFrog, GitHub blog + changelog, coding-agent changelogs
(Claude Code, Codex, Copilot, Cursor). All 8 added URLs fetched this session, HTTP 200,
title/ID present in body; all 7 arXiv items checked as v1-only (no version history).

+added:    Microsoft Research threat research: editing fewer than 10 neurons injects a sleeper-agent backdoor that passes evaluations and runs an attacker's command on a trigger string in context; inference-level compute is enough; Claude Code with Opus 4.5 found most of the threat autonomously; no in-the-wild weaponization observed — https://www.microsoft.com/en-us/security/blog/2026/11/10/detecting-sleeper-agents-hiding-in-ai-model-neurons/
+added:    arXiv 2611.13491 SURA: data-efficient unlearning-based backdoor detection for federated learning, reusing one small clean proxy across loss-ascent probes, restoring BN statistics — https://arxiv.org/abs/2611.13491
+added:    arXiv 2611.13577 TERM: target-free inference-time jailbreak defense guiding residual modulation by prompt-calibrated attention entropy — https://arxiv.org/abs/2611.13577
+added:    arXiv 2611.13684 BypassGym: first benchmark for AI agents building loaders that run C2 shellcode past live endpoint EDR on real machines — https://arxiv.org/abs/2611.13684
+added:    arXiv 2611.13960 ActiveSecHarness: coding-harness prompt injection turns bash/curl/find into Living-off-the-Land payloads; neurosymbolic reference monitor with HMAC capability tokens and four authority rings — https://arxiv.org/abs/2611.13960
+added:    arXiv 2611.14083 Sable: single-GPU argument system proving bit-exact token-for-token inference of a committed integer-quantized transformer (demo: Qwen2.5-32B, 64 layers, one RTX PRO 6000) — https://arxiv.org/abs/2611.14083
+added:    arXiv 2611.14143 GARD: target- and trigger-agnostic post-hoc backdoor defense for CLIP-style encoders using structure-matched adversarial representation probes — https://arxiv.org/abs/2611.14143
+added:    arXiv 2611.13367: impossibility theorem — no pair of complete machine-checkable allow/deny rulebooks can gatekeep which programs an AI agent may run — https://arxiv.org/abs/2611.13367

-rejected: arXiv keyword-list re-appearances with old v1 dates on today's announcement lists (2611.11399 RoCoMark v1 Nov 8/v2 Nov 10, 2504.18598 BadMoE v1 Apr 2025/v3, 2512.08289, 2603.16382, 2605.14786, 2602.08874, 2608.20047, 2610.02418) — replacement announcements, not first announcements, v1 outside 72h (sweep #5 rule; 2611.11399 additionally is code-watermarking/provenance research, not AI security)
-rejected: cs.CL keyword noise on today's list (TimeOmni-o1, Brasilidade, knowledge-editing capacity allocation, BioMem agentic memory, Mobile AI Factory, IFRS ensemble, EngineAgent, AutoScaffold and ~20 more) — general ML/NLP research with no security disclosure; keyword match on "LLM/agent/model" only
-rejected: arXiv 2611.13531 "Framing-Induced Vulnerability in Large Language Models" — the "vulnerability" is susceptibility to news-headline framing (social-science measurement), not an exploitable security flaw
-rejected: GitHub advisories Nov 11 02:40->12:00 window, all 137 rows (hMailServer ~20-bug disclosure, Wordfence/WordPress plugin bulk at 06:32, Go toolchain cgo build-time injection trio, Octopus Server trio, ManageEngine ADAudit Plus pair, keycloak, Padloc, Semantic MediaWiki batch, Directus authz bypass, AVideo auth bypass, Jellyfin path traversal, Akaunting XSS, MQTTX XSS) — generic platform/web/dev-tool security, no AI factor (spec section 2 scope)
-rejected: HN >=150 pts in window (2 stories): "Protecting SSH Keys with macOS Secure Enclave" (230) — generic security how-to, no AI factor; "A great old-timey game-programming hack" (189) — not security
-rejected: Techmeme window: Alphabet/Spirit Airlines data-acquisition privacy filing (regulatory consent decree), EU Digital Fairness Act (platform regulation, not a security disclosure), Nvidia/Megaspeed spy scandal + H200 export-control reporting (espionage/geopolitics, not an AI product security event), SoftBank AI turbine reservation, UK under-16 ban expansion, AI grading essay — no AI product-security angle
-rejected: BleepingComputer Nov 11 non-AI: Microsoft audio + Exchange sign-in update fixes, Cisco BroadWorks flaw actively exploited per CISA — generic platform security, no AI factor
-rejected: BleepingComputer "Railway breach leads to $740,000 in fraudulent AI charges" (Nov 10 17:20) — press coverage of the incident already added in sweep #6 from Railway's own report; no new disclosure facts
-rejected: OpenAI Nov 11 posts: ChatGPT Plus for veterans and servicemembers (08:00), Gilbert + Tobin customer story (Nov 12 00:00) — product/business posts, no security content
-rejected: Anthropic /claude-fable-and-mythos-5-2 page (sitemap lastmod Nov 10 21:27) — product model-page update, no security content; IPO roadshow page dispositioned in sweep #6
-rejected: Hugging Face blog Nov 11: BotPsy sub-billion-parameter on-device assistant — model/product post, no security content
-rejected: Coding-agent changelogs: Codex Nov 11 releases (rust-v0.167.0-alpha.4.2 at 01:29, alpha.10 at 03:33, alpha.11 at 10:11; bodies are a single "Release ..." line) — no changelog, no security entries (same handling as sweep #6); Claude Code — nothing newer than v2.1.322 (dispositioned in sweep #6); Cursor changelog 2026.11.10 ("Improvements to Cursor in Agents", "Redesigned hooks and improved dashboard") — product changes, no security fixes; GitHub changelog — nothing newer than the Nov 11 00:10 Copilot seat-management entry (dispositioned in sweep #6)

Swept, zero new items in window: Google Project Zero (latest Oct 27), Unit 42 (latest Nov 6,
rejected in #4), Embrace The Red (feed now reachable, latest Oct 13 VS Copilot RCE, older than
72h), AWS Security Blog (nothing newer than the GuardDuty post rejected in #6), Meta newsroom
(Nov 9, rejected in #4), Microsoft Security Blog (the sleeper-agent post above is its only new
entry), GitHub blog (Nov 10 Git security post rejected in #6), per-repo advisory endpoints
llama.cpp / transformers / LangChain / CrewAI / vLLM / PyTorch (nothing published since the
Nov 9-10 entries swept in #6), ReversingLabs (rss.xml still serves the Next.js HTML shell; no
post newer than the Oct 27 one noted in sweep #5), JFrog blog index (zero server-side article
links — Q3b open).

Method notes (not policy questions):
- GitHub global advisories: >=02:40 query caps again at 100 rows (newest 09:31:30); the
  two-sided range syntax published=A..B (comma form returns 422) enumerated the remaining
  02:40-06:32:38 slice in 48 rows; 137 unique rows in the window, keyword-scanned, zero
  AI-related.
- Embrace The Red: /blog/rss.xml returns the HTML app shell; the real feed is
  /blog/index.xml. Corrected feed path used, latest post Oct 13.
- Daily Swig: worse than sweep #6 reported — dailyswig.com and www.dailyswig.com both fail with
  SSL_ERROR_BAD_CERT_DOMAIN in a real browser (TLS certificate does not match the domain, so no
  HTTPS client can reach the site), and portswigger.net/daily-swig/ai 301-redirects to the
  PortSwigger homepage. Source currently unreachable; Q1 in data/questions.md stays open.
- Cursor changelog is unreachable to curl (JS shell for /changelog, .atom and .rss return HTML);
  read via the invisible-dots browser this sweep. Cursor was the only URL that needed it —
  every other source and all 8 added URLs answered 200 to direct fetch.
- NVD keyword searches Nov 11 02:40 -> 12:00: "prompt injection" 0, "LLM" 0, "AI agent" 0,
  "large language model" 0, "jailbreak" 0. The morning import batch was dispositioned in sweep #6.
- arXiv: Wednesday Nov 11 batch (sweep #6 hand-off) confirmed in RSS; 7 security-relevant cs.CR
  hits, each abs page v1-only (version history checked individually), submitted Nov 9-10,
  HTTP 200, title in body. Cross-listed old-ID entries dispositioned as replacements above.

## 2026-11-12 00:00 UTC — sweep #8

Window: Nov 11 12:00 UTC -> Nov 12 00:00 UTC (picks up where sweep #7 stopped). Sources swept:
OpenAI news (RSS), Anthropic (sitemap), Google Project Zero (feed), Microsoft Security Blog,
Meta newsroom + engineering, AWS Security Blog, Unit 42, Embrace The Red, Daily Swig (re-probed,
see method notes), BleepingComputer, arXiv cs.CR+cs.CL RSS, GitHub global advisories API (window
fully enumerated, see method notes) + per-repo advisory pages (llama.cpp, vLLM, transformers,
LangChain, CrewAI, PyTorch), NVD keyword searches, Hacker News >=150 pts (Algolia API), Techmeme
(feed.xml), Hugging Face blog, ReversingLabs, JFrog, GitHub blog + changelog + Copilot changelog,
coding-agent changelogs (Claude Code, Codex, Copilot, Cursor). All 4 added URLs and all 3
corroboration URLs fetched this session, HTTP 200, title/ID present in body.

+added:    Notion security advisory: the Fireworks AI vendor breach exposed Notion AI chat transcripts from the Insights dataset in a small number of workspaces; Notion cut vendor traffic and deleted its data by Nov 3; transcripts unattributable, no tool-call results, fewer than 1% of workspaces affected — this is the victim's own disclosure with new facts (which data, scope, mitigation), distinct from the sweep #6 rejection of the initial "Notion is a customer" statement that carried no new facts — https://www.notion.com/security/advisories/security-incident-involving-a-third-party-ai-vendor-nov9
+added:    Arctic Wolf Labs pack alert: attackers self-host AntiFox (Chinese-language open-source agentic pentesting platform, fork of ARTEX 0.1.1) on Alibaba Cloud against internet-facing web services and AWS environments; confirmed intrusions deployed an ASPX web shell on an exposed IIS server, parsed web.config credentials, burst-enumerated AWS services with valid IAM identities, read a production Secrets Manager secret, and read/copied/deleted S3 objects — https://github.com/rtkwlf/wolf-tools/tree/main/pack_alerts/202611-antifox-alibaba-hosted-platforms-facilitate-web-cloud-intrusions
+added:    GHSA-4f34-q9vf-fwfg (moderate, CVE-2026-126767): OpsTree REMS through 1.0.0 RCA chat assistant discloses its hidden system prompt and internal orchestration instructions via crafted chat input to POST /api/v1/rca/chat — https://github.com/advisories/GHSA-4f34-q9vf-fwfg
+added:    Microsoft FORGE Lab: codename MDASH investigated Linux kernel vulnerabilities, resulting in 59 upstream bug reports, 33 assigned CVEs and 38 local root privilege-escalation exploits; the write-up covers turning AI-generated findings into validated upstream fixes (new results from the same research line as the sweep #5 Hyenas post) — https://www.microsoft.com/en-us/security/blog/2026-11-11/lessons-from-59-linux-kernel-bug-reports/

-rejected: arXiv — no new announcement batch in the window: RSS still served the Wednesday Nov 11 list dispositioned in sweep #7; the Thursday Nov 12 batch announces after this window closes (hand-off note below)
-rejected: GitHub global advisories Nov 11 12:00->Nov 12 00:00 window, 265 of 271 unique rows after keyword triage (Checkmk pair incl. critical, Revive Adserver stored/reflected XSS batch, Kendo UI for jQuery pair, Dell Ansible/OpenManage batch, Apache Batik/FOP batch, BusyBox trio, Keycloak, Node.js https.Agent + file-watcher pair, WordPress plugin batch, AVideo, SiYuan path traversal, Spatie Browsershot local file read, social-auth-core login CSRF, Cockpit WebSocket DoS, AsiaInfo BSS XSS, NextScan hardcoded creds (keyword noise: "AuthAgent" is a scan-node router), proxygen/mvfst HTTP/3 pair, ByteStash session expiration, mvfst) — generic platform/web/dev-tool security, no AI factor (spec section 2 scope)
-rejected: OpsTree REMS sibling advisories published in the same Nov 11 21:32 batch (auth bypass, missing authentication, SSRF, stored XSS, improper authorization, improper attribution) — generic web flaws in a scheduling product; only the LLM chat-assistant prompt injection is an AI-security disclosure and entered the feed
-rejected: Helio CVE-2026-65957 / GHSA-2c3r-q7gv-hp2m (MCP proxy dashboard CORS private-prefix bypass) — GHSA published 2026-07-03 per the advisory page; only the NVD import (Nov 11 21:16) is new, same age rule as pgEdge in sweep #6
-rejected: taOS CVE-2026-67547 + CVE-2026-67548 / GHSA-h24f-gp4c-8qjm + GHSA-47g9-fwwp-hrfp (self-hosted AI agent OS: skill-exec endpoint and settings-route flaws) — GHSAs published 2026-07-04 per the advisory pages; only the NVD imports (Nov 11 22:17) are new, same age rule
-rejected: Unit 42 "ARTEX and AI-Powered Red Teaming Tools: What Defenders Need to Know" (Nov 11 23:00) — defensive-guidance synthesis of the ARTEX incident already carried in the feed (WorkOS, Railway); its new observations (IR engagements, a managed-hunting campaign) ship without verifiable specifics or IoCs; not a first disclosure
-rejected: Unit 42 "The XG Files: Falcon C2 Flying Under The (Browser) Radar" — browser-JavaScript C2 malware research, no AI factor
-rejected: Microsoft Security Blog "The global socket: how threat actors are weaponizing GSocket for stealthy C2" — generic C2 tradecraft, no AI factor
-rejected: BleepingComputer Nov 11 non-AI: Baker Tilly third-party breach, ScreenConnect MSP intrusions, eMudhra CA revocations, Maersk/WhiteArmor ransomware, eBPF rootkit trend piece, N-able zero-day hotfixes — generic platform security, no AI factor
-rejected: OpenAI RSS Gilbert + Tobin customer story (Nov 12 00:00) — business post, already dispositioned in sweep #7
-rejected: Anthropic sitemap /claude-fable-and-mythos-5-2 (lastmod Nov 11 18:10) — product model-page update, same handling as sweeps #6/#7
-rejected: Techmeme window (15 items): Nvidia board/Huang stories, CoreWeave loan, Lutnick/Megaspeed export-control dispute, OKXICE launch, Applied Intuition raise, Meta infra exec departure, Imprint raise, Edelbyte stealth launch, IBM Quantum Nighthawk, Gemini Decisions API (product), United Airlines AI-agent booking ban (ToS/policy, not a security disclosure), AWS Emulator, essays — no AI product-security disclosures
-rejected: HN >=150 pts window (16 stories), zero AI-security candidates: Maya site mapping with agents (911), AWS emulator, Jane Street smoke-alarm incident, TV ACR privacy, Nvidia, WebGL sand, United AI-agent booking ban (282), Apple forum closure, weight-space LP, Gemini Decisions API, work essay, Atuin Threads (product, no disclosure), goat sacrifices, Firefox fingerprinting research (non-AI), YC game, Yogurt
-rejected: AWS Security Blog: "Ease of use and secure defaults in the new AWS experience" (console security UX) + "Part 4 - Cybersecurity journeys" (career essay) — no AI-security content
-rejected: Meta engineering "Intelligent Delivery: Making Every Unit of Compute Serve Better Ads" — product/infra post. GitHub changelog "Issue types and fields are available for personal accounts" — not security; Copilot changelog feed nothing newer than the Nov 11 00:10 seat-management entry (dispositioned in sweep #6)
-rejected: Coding-agent changelogs: Codex window releases (0.170.0-alpha.1 20:09, 0.169.0-alpha.11.1 22:27, 0.167.2 22:46, alpha churn; the 0.167.2 body is a Vertex AI incremental-tools feature entry only) — no security entries; Claude Code — nothing newer than v2.1.322 (dispositioned in sweep #7); Cursor changelog read via the MCP browser this sweep — newest entry Nov 6 "Voice Chat with Agents", product feature, no security fixes, nothing in window

Swept, zero new items in window: Google Project Zero (latest Oct 27), Embrace The Red (latest
Oct 13), Hugging Face blog (latest Nov 11 05:01 BotPsy, dispositioned in #7), Meta newsroom
(Nov 9), GitHub blog (Nov 10), per-repo advisory pages llama.cpp / transformers / LangChain /
CrewAI / vLLM / PyTorch (nothing newer than the PyTorch Nov 10 GHSA dispositioned in #6),
Daily Swig (re-probed, still dead, see method notes), ReversingLabs (rss.xml still serves the
Next.js HTML shell; Q9 open) and JFrog blog index (still zero server-side article links; Q3b
open).

Method notes:
- GitHub global advisories: range queries re-return the same newest 100 rows when paged; the
  window was enumerated via nested published= slices (12:00..15:31:01 = 99 rows below the cap,
  15:31:01..18:31:38, 18:31:38..21:33:02 = 100 rows; nothing published after 21:33:02),
  271 unique rows, keyword-scanned. Per-repo advisory API answers 404 unauthenticated; the
  advisory HTML pages were read instead.
- NVD keyword searches Nov 11 12:00 -> Nov 12 00:00: "prompt injection" 1 (OpsTree REMS, added),
  "AI agent" 3 (Helio + taOS pair, age-rejected above), "LLM" 0, "jailbreak" 0,
  "large language model" 0.
- Microsoft's kernel-research post serves curl only the page metadata for its body (JS-rendered
  article); all added facts (MDASH, 59 reports, 33 CVEs, 38 exploits, AI-generated findings)
  were present in the fetched page's embedded description. Direct MCP-browser navigation to
  microsoft.com hung at TLS commit; the browser proxy refused connections this sweep and the
  browser was opened with direct egress, sufficient for the Cursor changelog.
- Daily Swig: re-probed this session, SSL cert-domain mismatch persists on the apex; Q1 stays
  open.
- arXiv: the Thursday Nov 12 announcement lands after this window closes; the RSS still carried
  the dispositioned Wednesday batch at sweep time — next sweep picks it up (same hand-off
  pattern as #6 -> #7).
## 2026-11-12 12:00 UTC — sweep #9

Window: Nov 12 00:00 UTC -> Nov 12 12:02 UTC (picks up where sweep #8 stopped; includes the
Thursday Nov 12 arXiv announcement sweep #8 handed off). Sources swept: OpenAI news (RSS),
Anthropic (sitemap), Google Project Zero (feed), Microsoft Security Blog + Microsoft Research
blog (Research blog added this round, see Q10), Meta newsroom, AWS Security Blog, Unit 42,
Embrace The Red, Daily Swig, BleepingComputer, arXiv cs.CR+cs.CL (RSS + /list/recent
cross-check), GitHub global advisories API (window fully enumerated, 287 unique rows, see
method notes) + per-repo advisory pages (llama.cpp, vLLM, transformers, LangChain, CrewAI,
PyTorch), NVD keyword searches, Hacker News (>=150 pts plus a keyword scan over all 365
in-window stories), Techmeme (feed.xml), Hugging Face blog, ReversingLabs, JFrog Research,
GitHub blog + changelog + Copilot changelog, coding-agent changelogs (Claude Code, Codex,
Copilot, Cursor). The added item's URL was fetched this session, HTTP 200, GHSA id + CVE id +
product name all present in body; every rejected candidate URL was also fetched this session.
invisible-dots was needed for exactly two URLs this round (Cursor changelog, SpaceBattles
thread), both for rejects.

+added:    Moonshot AI Kimi Code workspace-scope security decision relies on untrusted additional_dir input (GHSA-75pq-p477-6h9c, CVE-2026-126758, CVSS 4.4, exploit public, fixed 2.1.0, vendor unresponsive to disclosure per the advisory) — https://github.com/advisories/GHSA-75pq-p477-6h9c

-rejected: Kimi Code sibling advisory GHSA-wpx2-j9xr-rxgh (CVE-2026-125757, open redirect via isOriginAllowed in the bundled Web UI) — same batch, same vendor; generic web-surface bug that never touches the agent's trust boundary, same treatment as the tsParticles MCP and Dagu rejections (spec section 2 scope)
-rejected: "Humble Bundle just killed most of their security through vibecoding" (HN 301, forum thread) — report is from an anonymous forum account, no named researcher, no PoC/repo/CVE, no vendor statement; the thread itself returns HTTP 403 status even to a real browser, and the only "second source" it cites is an anonymous Reddit thread; claims (2FA removed, no password confirmation on email/password change) cannot be pinned to a primary artifact (spec section 3)
-rejected: Microsoft Research "Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents" (Nov 10 20:52, HN 1 pt) — the Research-blog write-up of the same fewer-than-10-dormant-neurons study already added from Microsoft's Security Blog in sweep #7 (same finding, same Claude-Code-with-Opus-4.5 origin); no new disclosure facts; source-coverage gap recorded as Q10
-rejected: Tell HN "Gemini Hydrogen Reachable via Vertex" (2 pts) — one-line anonymous claim that a model is reachable on a public endpoint; no evidence, no named researcher, no corroboration (spec section 3)
-rejected: arXiv — the Thursday Nov 12 announcement batch had not landed at sweep time: the RSS channel was empty (lastBuildDate 05:00 UTC) and /list/cs.CR/recent + /list/cs.CL/recent both still top out at the Wednesday Nov 11 group dispositioned in sweep #7; hand-off to the next sweep (same pattern as #6 -> #7)
-rejected: GitHub global advisories Nov 12 00:00->12:02 window, 287 unique rows (00:32:55-00:33:04 NVD-import burst of ~102 rows plus the 09:32 batch) — AI-keyword scan surfaced only the two Kimi Code advisories (one added above, one rejected) and Directorist GHSA-qvhr-hcf3-738r ("AI-Powered Business Directory" WordPress plugin; generic web bug, AI only in the product name); rest generic platform/web/dev-tool security (spec section 2 scope)
-rejected: Reasonix maintenance-stop announcement (esengine/DeepSeek-Reasonix discussion 15110, "will stop being maintained" from Nov 12, HN 1 pt) — project-lifecycle news, not a security disclosure (MS Research AI Frontiers shutdown precedent, sweep #4); the project's CVE-2026-123653 prompt-injection item is already in the feed from sweep #1
-rejected: OpenAI Nov 12 posts: "Supporting research on AI and children's safety" ($2M research commit), Holiday Extras Codex customer story, Gilbert + Tobin customer story — policy/society/business posts, no security disclosure
-rejected: Meta "New Agreements Bring Clean Manufacturing to Data Center Supply Chains" (Nov 12 12:00) — ESG/supply-chain procurement, not security
-rejected: BleepingComputer window: Windows KB5132061 update-install errors, "Hackers plant GSocket on Linux servers" (same generic-C2 line as the Microsoft GSocket post rejected in sweep #8) — generic platform security, no AI factor
-rejected: Techmeme OpenAI item: Russian pranksters tricked three OpenAI natsec/policy executives into a video call posing as Ukrainian officials (also The Intercept, HN 7) — social engineering against company staff, not an AI-product security flaw (Nvidia spy-plot precedent, sweep #2)
-rejected: Techmeme window remainder: Grok image-abuse lawsuits against xAI (legal), Lumbra $25M for agent orchestration in secure enclaves (funding), "handoverism" essay, Nvidia/Huang scandal items, SMIC/Foxconn results, Tencent speech, robot raises — corporate/legal/finance, no security disclosure (spec section 3)
-rejected: HN "Critical RCE Vulnerability in Apache HTTP Server" (243) — generic web-server bug, no AI factor
-rejected: HN in-window keyword-scan remainder (365 stories): AI-bubble/IPO/opinion posts, product launches (DeepL Aure, MAI-Transcribe-2, Claris Harbor), Show HN tools shipped without any security disclosure (Taster prompt-rule guardrail, bla anti-reverse-engineering obfuscator, swapback PII proxy, ShanHaiGateway), a user-jail cautionary tale, OpenAI status.openai.com workspace-agent availability blip (outage, not a security incident) — outside the bar
-rejected: Coding-agent changelogs: Codex 0.170.0-alpha.2 (Nov 12 01:34, body is a single "Release ..." line, same handling as #6-#8), Claude Code (nothing newer than v2.1.322, dispositioned in #6), Cursor changelog read via the MCP browser (newest entry still Nov 6 "Voice Chat with Agents", product, no security fixes), Copilot changelog and GitHub changelog (nothing newer than the Nov 11 entries dispositioned in #8) — no security entries this window
-rejected: JFrog Research posts reachable again: "Contagious Interview - One Interview Away From Infection" (Nov 9; >72h and generic developer-targeted malware where the AI coding assistant is only the victim-side detector), "Beyond testmetest: 42 Armed GitHub Repositories" (Oct 26, >72h), "Byte Debt" Vert.x header typo (Oct 28, >72h, generic web bug) — age rule (spec section 3), no backfill

Swept, zero new items in window: Google Project Zero (feed parsed, nothing Nov 12), Microsoft
Security Blog (latest Nov 11 19:49, dispositioned in #8), AWS Security Blog (latest Nov 11
19:57, dispositioned in #8), Unit 42 (latest Nov 11 23:00 ARTEX synthesis, dispositioned in
#8), Embrace The Red (latest Oct 13), Meta newsroom (only the clean-manufacturing post above),
Hugging Face blog (latest Nov 11 05:01 BotPsy, dispositioned in #7), GitHub blog (Nov 10),
per-repo advisory pages llama.cpp / transformers / LangChain / CrewAI / vLLM / PyTorch (newest
listed is the PyTorch Nov 10 GHSA added in #6; the two vLLM Nov 9 unreviewed advisories added
in #6 no longer appear on the repo list page — direct URLs untouched), ReversingLabs (rss.xml
still a Next.js HTML shell, Q9 open), Daily Swig (still dead, Q1 open).

Method notes (not policy questions):
- GitHub global advisories cap again at 100 rows/query and page= still re-returns page 1. The
  window hid a bulk NVD-import burst at 00:32:55-00:33:04 UTC (>=100 rows in 10 seconds):
  enumerated by nesting published= slices down to one-second ranges (17/16/14/12/1/4/15/23
  rows) plus a 44-row lower slice and the 41-row 00:33:05-06 boundary slice; 287 unique rows,
  all keyword-scanned. The burst moment is why dedup against data/items.json stays in triage.
- NVD keyword searches Nov 12 00:00 -> 12:02: "prompt injection" 0, "LLM" 0, "AI agent" 0,
  "jailbreak" 0, "large language model" 0.
- JFrog: research.jfrog.com serves server-side article links to plain curl (the JS-shell
  finding from #1-#8 was www.jfrog.com/blog) — swept all three current posts above; the Q3b
  gap is at least partly closed by using research.jfrog.com as the JFrog entry point.
- vLLM repo advisory list page no longer lists GHSA-v97p-wpxp-px84 / GHSA-fpch-3xrf-7q96
  (added in #6); their direct pages are still the primary sources recorded in items.json.
  Recorded in case a future sweep treats their disappearance as a withdrawal signal.
- invisible-dots used for exactly two URLs (Cursor changelog JS shell, SpaceBattles 403);
  the container was closed and removed after the sweep per operator instruction.
## 2026-11-13 00:45 UTC — sweep #10

Window: Nov 12 12:02 UTC -> Nov 13 00:52 UTC (picks up where sweep #9 stopped; the pending
Thursday Nov 12 arXiv batch again had not landed at sweep time, see below). Sources swept:
OpenAI news (RSS), Anthropic (sitemap), Google Project Zero (feed), Microsoft Security Blog +
Microsoft Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red, Daily Swig,
BleepingComputer, arXiv cs.CR+cs.CL (RSS + /list/recent cross-check), GitHub global advisories
API (window fully enumerated in slices, 327 unique rows + 1 tail row, see method notes) +
per-repo advisory pages (llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch), NVD keyword
searches, Hacker News (>=150 pts plus a keyword scan over all 763 in-window stories), Techmeme
(feed.xml), Hugging Face blog, ReversingLabs, JFrog Research (research.jfrog.com + its RSS),
GitHub blog + changelog + Copilot changelog, coding-agent changelogs (Claude Code, Codex,
Copilot, Cursor). All nine added items' URLs were fetched this session, HTTP 200 (Apify page +
its statuspage JSON API, the seven vendor/advisory pages direct), with the product/CVE/GHSA
names present in body; every rejected candidate URL was also fetched this session.
invisible-dots was needed for exactly three URLs this round (CloudSEK report behind a Cloudflare
challenge, Cursor changelog, OpenAI cyber page) — see method notes.

+added:    Microsoft "Trained to betray" — MS Security Research + AI Red Team studies: dataset poisoning teaches coding agents to discover/misuse tools, model poisoning builds backdoors that persist through later post-training and transfer; mitigation guidance included — https://www.microsoft.com/en-us/security/blog/2026-11-12/trained-to-betray-how-post-training-supply-chain-attacks-hijack-coding-agents/
+added:    Unit 42 CL-STA-1238 — attackers ran a forked Cairn autonomous-agent framework (AI-driven agent workers, shared blackboard, human hints) against semiconductor, defense and civil-society sectors; controller + egress infrastructure documented — https://unit42.paloaltonetworks.com/cairn-framework-ai-autonomous-attacks/
+added:    CloudSEK "Off the Rails" — exposed Silent Lynx server shows a Kazakhstan-linked espionage op using the open-source Hermes AI agent (mostly DeepSeek models) for recon, data collection and evidence cleanup; Tajik railway, Uzbek forestry, defence documents — https://www.cloudsek.com/blog/off-the-rails-inside-an-exposed-silent-lynx-server (corroborated: https://therecord.media/kazakhstan-hackers-central-asia)
+added:    Perplexity "Escaping SPACE: Part II" — nine-model red-team rerun: no VM-host escapes, but Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol and Kimi K3 reached blocked destinations through approved partial egress (PyPI, Ubuntu mirrors); fixed and re-verified — https://www.perplexity.ai/hub/blog/escaping-space-part-ii
+added:    OpenAI Alignment — named-team paper on improving "fuzzy robustness" of LLM safety in dual-use cyber/biology domains via automated red teaming — https://alignment.openai.com/fuzzy-robustness/
+added:    Opera Neon MCP Hub CVE-2026-123139 — malicious browser extension can send commands through the MCP Hub to the built-in Browser Operator extension and take over the browser (Opera advisory, Nov 12) — https://security.opera.com/en/improper-verification-of-source-of-a-communication-channel-opera-security-advisories
+added:    Apify security incident — compromised credential used Nov 9-10 to harvest account credentials and access customer accounts; Nov 12 update confirms customer personal information/stored data was accessed, notifications going out, OAuth connections revoked, still investigating — https://status.apify.com/incidents/sxd1dj5j6749
+added:    pgEdge RAG Server GHSA-w437-p477 note: shared BM25 keyword index leaked one tenant's documents to another under ordinary concurrent queries, defeating tenant scoping (CVE-2026-125325, <=1.0.0) — https://github.com/advisories/GHSA-w437-9mwm-j6vq
+added:    pgEdge LLM credential disclosure — RAG Server relayed provider error text (with partially masked API-key fragments) to unauthenticated callers; sibling Go LLM library health endpoint does the same with its real key (CVE-2026-125326, corroboration GHSA-8835/CVE-2026-125328) — https://github.com/advisories/GHSA-68g8-gh8f-7rhg

-rejected: pgEdge RAG Server GHSA-5qg5-rgp6-j5m5 (CVE-2026-125327, same vendor batch) — unauthenticated resource exhaustion via unbounded corpus read; availability-only DoS, no data or trust-boundary breach, the two disclosure bugs carry the story
-rejected: SakanaAI text-to-LoRA GHSA-25cm-vgcw-2gfw (CVE-2026-126996) — VulDB-sourced NVD import, low severity, boilerplate "argument manipulation" path traversal in a research demo repo's checkpoint loader; vendor-unresponsive research code without a product surface (scope handling as the Kimi sibling advisories in #9)
-rejected: GitHub global advisories Nov 12 12:02->00:52 remainder, 327 unique rows enumerated (plus the single 23:46-00:52 tail row) — the 18:31-23:46 stretch is dominated by NVD-import batches: Rails Action View/Pack XSS + DoS set (incl. GHSA-f3w6 critical arbitrary file write/read in Active Storage), Loofah scrubber bypass set, Hono/Astro/Addressable/AppFlowy, XWEB EVO, AWX debug endpoints; the one critical outside those, GHSA-h94v-vx95-5jpx (ScreenConnect RMM+ extension static default API secret), is a generic RMM product bug with no AI factor (same product family as the SCKit worm coverage, still out of scope); datamodel-code-generator is data-model codegen, not ML; Directorist GHSA-qvhr keeps failing on "AI" only being in the product name (re-published in window, disposition unchanged from #9)
-rejected: BleepingComputer "Hackers compromise 4,800 GitHub accounts to steal cloud credentials" (wfx-loot / Aikido research) — serious CI/CD secret-theft campaign (4,837 accounts, malicious GitHub Actions workflows, ~50 accounts with package-publishing rights), but zero AI factor in the mechanism; same bar as the Apache HTTP/2 rejection (spec section 2 scope)
-rejected: Techmeme/FBI "FaceStealer" item (Reuters) — FBI says North Korean hackers stole $400M+ in crypto over 18 months via fake video calls with AI deepfakes; large-scale AI-enabled fraud, but no primary artifact yet (law-enforcement statement to press, no advisory/report), and deepfakes here are human-targeted social engineering (Nvidia spy-plot precedent, sweep #2); will pick it up if a primary advisory lands
-rejected: BleepingComputer remainder in window: worm-like SCKit (ScreenConnect), N-able "4EverAble2Shell" zero-day, Apache HTTP/2 RCE patch news (same story as the HN-243 reject in #9), Windows SignDrop privilege zero-day, Gootloader/CastleRAT/StealC takedown, "your endpoint management platform" sponsored post — generic platform security or marketing, no AI factor; Massive Microsoft 365 outage affecting Outlook/Teams/Copilot — availability incident, not a security disclosure
-rejected: OpenAI in window: "Cyber at the Frontier" hub + "Daybreak Red Team" access form (HN 5/3) — program/product launch (agent access applications for cyber defense), no disclosure; the hub page 403s to curl and was read in the browser (see method notes); Amplitude + Gondola Codex customer stories — business posts (#9 handling of the same feed continues)
-rejected: Anthropic — sitemap "s1-sec" page is the Oct 26 IPO S-1 filing announcement (funding, not security); "news/claude-4" sitemap lastmod bump points at the May 2025 Opus 4 / Sonnet 4 announcement, a re-render, not a new disclosure
-rejected: GitHub blog Nov 12 "Build once, run on any agent: a practical guide to agent skills" — how-to/product; GitHub changelog Nov 12 items (Actions ARM32 retirement, Secret scanning detects Cloudflare service tokens, public monitoring alert links) — platform capability and housekeeping, not security disclosures; Copilot changelog nothing newer than the Nov 11 entries dispositioned in #9
-rejected: Microsoft Research "Microsoft Research Focus — Edition #98" (Nov 12 podcast roundup) — podcast digest, no new disclosure (Lazare/Research-blog precedent, sweep #9)
-rejected: Hugging Face blog "LLMON: An LLM-Native Markup Language, now in Granite 4.1" (Nov 12 19:07) — model/product release post, no security content
-rejected: Coding-agent changelogs: Claude Code v2.1.323 (Nov 12 22:12 — adds opt-in Code Mode, plugin version-range enforcement, Bash glob-variable warning; features/hardening, no disclosed vulnerability fixed, spec section 2 changelog rule) and Codex rust-v0.168.0 plus three Nov 12 alphas (release bodies are single "Release ..." lines, same handling as #6-#9); Cursor changelog re-read in the browser — newest entry still Nov 6 "Voice Chat with Agents"
-rejected: arXiv — the Thursday Nov 12 announcement batch still had not landed at sweep time (third consecutive carry): RSS lastBuildDate Nov 12 05:00 UTC, both cs.CR/cs.CL channels empty, /list/recent still topped by the Wednesday Nov 11 group; hand off to sweep #11
-rejected: JFrog Research "bREACHYing the Perimeter: How One HTTP Request can Hijack your AI Robot" — in-scope (Reachy Mini AI robot, single-HTTP-request takeover) but published Nov 3, 10 days old, first surfaced by us; age rule (spec section 3) and the no-backfill precedent from #9 keep it out — policy question appended as Q11; Contagious Interview (Nov 9) and the Oct posts in the research RSS remain >72h (re-confirms #9)
-rejected: HN >=150 remainder: Jensen Huang resignation stories (up to 2,774 pts; corporate/leadership news, not security), Rust 1.100 release (routine, no security note), M365/Azure outage threads (availability), Gemini 4 Hydrogen launch posts (product), embroidery-machine subscription, smolmaxxing, Nintendo DMCA emulator sweep (IP/legal), Zen 6 interview, dwarf planet, data-center power equity, Jane Street interviews, Quake 4 mod
-rejected: HN keyword-scan remainder (763 stories): Will DePue "Chinese frontier AI models carry the fingerprint of export-controlled GPUs" (96 pts, named researcher, primary report — export-control forensics, not a security disclosure), Apple Foundation Models framework + Libreagi + AgentRun + Foundation-models/decision-model posts (SDK/product/tools shipped without a disclosure), United Airlines agent-blocking op-ed, NYT agent-behaviour feature, anonymous claims and Show HN agents — outside the bar (spec section 3)

Swept, zero new items in window: Google Project Zero (latest Oct 27), AWS Security Blog (latest
Nov 11 19:57, dispositioned in #9), Microsoft Security Blog (only the added post, others
dispositioned in #8/#9), Unit 42 (only the added post; ARTEX synthesis + Falcon C2 dispositioned
in #9), Embrace The Red (latest Oct 13, feed now read via /blog/index.xml), Meta newsroom
(nothing newer than the clean-manufacturing post dispositioned in #9), per-repo advisory pages
llama.cpp / transformers / LangChain / CrewAI (nothing newer than Oct; page dates checked),
PyTorch (Nov 10 GHSA, added in #6) and vLLM (Nov 2; the two Nov 9 GHSAs still absent from the
repo list, #9's note stands), ReversingLabs (rss.xml still a Next.js HTML shell, Q9 open),
Daily Swig (domain does not resolve at all tonight, Q1 open), NVD keyword searches Nov 12
12:02 -> Nov 13 00:20: "prompt injection" 0, "LLM" 0, "AI agent" 0, "jailbreak" 0.

Method notes (not policy questions):
- GitHub global advisories window re-sliced into 2-hour queries (all <=100 rows, no paging
  needed); 327 unique rows keyword-scanned. One caveat observed this round: a two-parameter
  published= range query near the window end silently ignored the lower bound and returned
  newest-100; re-verified with a single-sided published=>=23:46 query, which returned exactly
  one row (Rails Active Storage, rejected above). The 00:32/09:32 UTC NVD burst pattern did not
  bite this window; dedup against data/items.json still runs in triage.
- Apify's status page is a JS shell to curl; incident text was read from the same page's
  Atlassian statuspage JSON API (incident sxd1dj5j6749, three updates; the "customers notified"
  update at Nov 12 16:16 UTC is the disclosure event inside the window). Source URL recorded is
  the human-readable incident page (200 via shortlink redirect).
- invisible-dots used for exactly three URLs: cloudsek.com (Cloudflare "Just a moment" challenge
  first, HTTP 200 + full report on the second navigation — primary source content confirmed:
  Silent Lynx, Hermes agent, DeepSeek models, Nov 11 2026), cursor.com/changelog (direct
  navigation to www.cursor.com/changelog timed out at domcontentloaded twice, 45s each;
  cursor.com/changelog loaded fine and was read — newest entry Nov 6, no security fixes), and
  openai.com/cyber-at-the-frontier (403 to curl, 200 in browser; program hub, rejected above).
- The container used for this round runs profile /state/profiles/sweep10 (a spare profile — the
  shared main profile was locked by another session's live browser; nothing else's container was
  touched mid-sweep). Per operator instruction all invisible-mcp-vnc containers are stopped and
  removed after the sweep.

## 2026-11-13 12:00 UTC — sweep #11

Window: Nov 13 00:52 UTC -> Nov 13 12:20 UTC (picks up where sweep #10 stopped at 00:52; the
carried Thursday/Friday arXiv announcement batch HAS landed this time - via /list/recent only,
the RSS channels are still empty upstream, see method notes). Sources swept: OpenAI news (RSS),
Anthropic (sitemap), Google Project Zero (feed), Microsoft Security Blog + Microsoft Research
blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red, Daily Swig, BleepingComputer,
arXiv cs.CR+cs.CL (RSS + /list/recent Fri-13 groups, 205 unique entries), GitHub global
advisories API (91 rows, whole window, every row eyeballed) + per-repo advisory pages (llama.cpp,
vLLM, transformers, LangChain, CrewAI, PyTorch), NVD keyword searches, Hacker News (>=150 pts
plus a keyword scan over all 360 in-window stories), Techmeme (feed.xml), Hugging Face blog
(incl. malware-scanner posts), ReversingLabs blog (browser + sitemap - Q9 gap now readable),
JFrog Research (browser - posts list last updated Nov 9), GitHub blog + changelog + Copilot
changelog, coding-agent changelogs (Claude Code, Codex, Copilot, Cursor). All eight added items'
URLs were fetched this session, HTTP 200, with the product/CVE/GHSA/paper names present in body;
each rejected candidate URL was fetched this session too (the one exception, the Computer Weekly
wind-farm report coverage, is itself a logged reject: 403 to the validator).

+added:    arXiv 2611.16693 MobileProbe — trajectory-aware environmental injection: instructions hidden in ordinary screen UI hijack mobile GUI agents mid-task, adaptive step selection beats fixed payload benchmarks (announced Fri Nov 13 batch) — https://arxiv.org/abs/2611.16693
+added:    arXiv 2611.14864 — fabricated risk warnings in low-trust content push tool-using agents into harmful over-abstention (availability attack); verify-evidence-authority defense — https://arxiv.org/abs/2611.14864
+added:    arXiv 2611.14808 OverServe — tool--prompt co-design makes subscription SaaS agents extend execution and amplify provider-paid model/tool cost (same pattern as the vLLM amplification GHSA added in #8) — https://arxiv.org/abs/2611.14808
+added:    arXiv 2611.17115 CapLite — across 15 open-weight models (13 at 1.7B-12B) the security artifacts that CaMeL/Progent/IPIGuard/DRIFT ask the agent's own model to write admit attacks; defense without model-written artifacts — https://arxiv.org/abs/2611.17115
+added:    arXiv 2611.14970 Poison-then-Hide — dormant backdoor implanted in a pretrained vision encoder wakes during full-model fine-tuning, where prior finetuning-activated attacks fail (distribution-time supply chain) — https://arxiv.org/abs/2611.14970
+added:    arXiv 2611.14950 — sandbaggers (small models failing GSM8K only under an audit marker) defeat detectors that were validated once against a fixed attack, once the model adapts — https://arxiv.org/abs/2611.14950
+added:    n8n GHSA-rc3w-8jc8-7x3w (CVE-2026-126386) — default-on, unauthenticated MCP OAuth dynamic client registration accepted an arbitrary redirect_uri; crafted authorization URLs redirected visitors before instance checks (no tokens in redirect); fixed 1.123.89/2.45.5/2.46.1/3.0.1, CVSS 5.1 — https://github.com/n8n-io/n8n/security/advisories/GHSA-rc3w-8jc8-7x3w
+added:    SiYuan GHSA-6pg3-pqxw-jm82 (CVE-2026-126393) — the MCP skill-install download path was left off the ssrfSafeClient pinning added by GHSA-x8gv-g2g3-65fj; redirect + DNS rebinding reach internal addresses on v3.8.8; fixed 3.8.10, CVSS 6.1 — https://github.com/siyuan-note/siyuan/security/advisories/GHSA-6pg3-pqxw-jm82

-rejected: Computer Weekly "AI agents bring down wind farm within two hours" (HN 1 pt) — Shinobi Security red-team result (Claude 66 min / Kimi ~2 h to shut a replicated 10 MW wind farm from VPN creds, head of research David Mound named) but coverage is 403 to the validator (Cloudflare), no second source surfaced, and no public Shinobi report found on shinobi.security (blog checked, homepage posts checked) — spec section 3: 403 without corroboration is rejected with reason
-rejected: Codex 0.169.0 stable (Nov 13 08:39, first stable inside the window) — Bug-Fix line "codex exec keeps project configuration and MCP servers disabled in untrusted directories, even with --sandbox workspace-write" is a trust-boundary hardening with no disclosed vulnerability, same handling as the Claude Code v2.1.323 hardening line in #10; alphas 0.170.0-alpha.4/0.171.0-alpha.1-3 bodies are single "Release ..." lines (#6-#10 handling); Claude Code changelog nothing newer than v2.1.323 (dispositioned in #10)
-rejected: GitHub global advisories Nov 13 00:52->12:00, 91 rows enumerated and eyeballed row by row (single page, asc order, window end re-verified against the max row 09:31:47) — batches: NetScaler "Memory overflow" set (~15), Dell PowerScale InsightIQ set (16), Edito CMS set (9), WordPress-plugin authz/XSS set (19), MCUboot serial-recovery, Rucio SSH impersonation, Linksys, TinyVNC/UltraVNC, Cairo-dock, SourceCodester, Kroxylicious ACL parser, Red Hat Quay reflected error, MISP Modules open redirect; zero rows with an AI factor (spec section 2 scope)
-rejected: HN >=150: Nvidia employees petition to bring back Jensen Huang (2,375+264 pts; corporate), M365 outage threads incl. CNN "Global Microsoft Outage" (availability, #10 M365-outage precedent), "Elevated errors on claude.ai, Claude Code and platform" status incident (240 pts; availability, fetched — incident text is elevated-errors only, no security disclosure), IceWM; keyword scan over all 360 stories: nothing >=150, sub-bar hits incl. Brave x402 agentic payments (131, product), Claude-down chatter, "Anthropic killed a Claude Code feature" complaint issue, DePue export-GPU reposts (#10 disposition unchanged), Notion/Fireworks (#10 added), Rust-std memory-safety post (no AI factor), Apache USN-9090-1 (#9/#10 Apache lineage, no AI factor), patent-suit + Maya-site + wind-farm items (above)
-rejected: Techmeme Nov 13 items: OpenAI Daybreak Red Team early-access + Daybreak Defense Factory (07:25) — same program launch rejected in #10 after reading the hub in the browser (no disclosure; #10 fetch stands); xFusion/Nscale IPO + Nvidia-backstop stories (funding), Samsung-Modine talks (business), Meta Muse outbound calling (product), Chinese compute-projects review (finance analysis), $2M teen-lawsuit payouts + Nvidia-spy-plot disagreements (legal/forensics, no security artifact)
-rejected: BleepingComputer in window: Check Point zero-day exploited for root access (Nov 13 10:44; firewall product vuln, no AI factor — Apache HTTP/2 precedent), 354 dark-web arrests (law-enforcement roundup, no AI factor), Windows activation failures after Nov updates (not a security disclosure)
-rejected: Anthropic sitemap new entry legal/self-hosted-managed-agents-license (Nov 12 17:16, license terms text, not a security disclosure; sibling claude.com "self-hosted Claude Managed Agents" HN post + "Securing Frontier AI" partner PR = program/legal pages)
-rejected: arXiv (Fri-13 batch remainder): 2611.15118 Lazare re-announcement — same dormant-neurons study added from the Security Blog in #7, research-blog duplicate rejected in #9 (Q10); 2611.17294 Metal Sandbox — coding-agent GPU isolation infrastructure where the abstract's focus is scheduling/utilization, not a disclosure; 2611.17123 CodeQL taint-tracking robustness and 2611.16565 decision models finding memory-safety bugs — AI-for-security, not security-of-AI (#7 keyword rule); 2611.15583 image-hiding model attacks, 2611.14929 speech-tokenizer semantic attacks, 2611.16151 fault injection on Q-learning, 2611.14594 adversarial-epsilon theory — outside the spec section 2 keyword list; 2611.17280 synthesis smuggling + 2611.16048 recurrent-state privacy leakage — privacy research out of niche (#7 Private-Vulnerability-Testing precedent); 2611.16183 safety alignment, 2611.14897 social-engineering detection, plus the non-AI set (FPGA anti-tamper sensors, 6G ISAC, power-grid co-sim, spectrum obfuscation, MPC data-value, IoT firmware bench, vuln-localization dataset) — same out-of-list handling as #7's batch

Swept, zero new items in window: OpenAI news (latest Nov 12 20:00, customer stories dispositioned in #10; nothing Nov 13), Google Project Zero (latest Oct 27), AWS Security Blog (latest
Nov 11 19:57, dispositioned in #9), Microsoft Security Blog (latest Nov 12, added in #10),
Microsoft Research blog (latest is the Focus #98 podcast rejected in #10), Meta newsroom (latest
Nov 12, dispositioned), Unit 42 (latest Nov 12, added in #10), Embrace The Red (latest Oct 13),
Hugging Face blog incl. malware-scanner posts (latest Nov 12 19:07 LLMON, rejected in #10),
GitHub blog (availability report Oct 2026 posted Nov 13 00:20, before the window and not a
security disclosure) + changelog + Copilot changelog (nothing newer than the Nov 12 items
dispositioned in #10), per-repo advisory pages llama.cpp / transformers / LangChain / CrewAI
(nothing newer than Oct; page dates checked) and vLLM (Nov 2; the two Nov 9 GHSAs still absent
from the repo list, #9's note stands) and PyTorch (Nov 10 GHSA, added in #6), Claude Code
changelog (nothing newer than #10), Cursor changelog re-read in the browser (newest entry still
Nov 6 "Voice Chat with Agents"), Daily Swig (domain still does not resolve, Q1 open).
ReversingLabs: the Q9 gap is now closed — the /blog listing renders in the browser and the
sitemap carries per-post lastmod; the newest posts are the two dated Nov 4 ("Securing Frontier
AI" partner post, "The AI Cyberattack Surge Hasn't Arrived"); their Nov 12 / Nov 10 sitemap
lastmod bumps are re-edits of those posts, not new disclosures (#10 claude-4 lastmod precedent),
so nothing was published inside the window; Binance-heist/pyenv and the other posts are >72h.
JFrog Research (browser): posts list "Last Updated On 9 Nov, 2026" — newest posts are the ones
dispositioned in #10 (bREACHYing Nov 3 rejected with Q11, Contagious Interview Nov 9); Byte Debt
(Vert.x session fixation), Beyond testmetest, Brotli2Shell, PHP2Shell, LiteLLM 1.107.0 hijack,
curl HTTP/2 RCE and GemStuffer are older and outside the window; nothing AI-new to disposition.
NVD keyword searches Nov 13 00:52 -> 12:00: "prompt injection" 0, "LLM" 0, "AI agent" 0,
"jailbreak" 0, "MCP" 2 (both added; NVD imported the pair at 10:17 UTC).

Method notes (not policy questions):
- arXiv: the Thursday announcement batch that sweeps #8-#10 carried has landed, but only on the
  /list/recent pages - the cs.CR and cs.CL RSS channels built today (lastBuildDate Nov 13 05:00)
  still contain zero items for the fourth consecutive round. Treated /list/recent as the source
  of truth: Fri-13 groups = 70 cs.CR + 135 cs.CL entries (205, 5 cross-listed deduped), all
  titles keyword-scanned, 12 candidates, 6 added, each abstract page fetched and title verified
  present in body.
- GitHub advisories: single asc-ordered published=>=00:52 query returned the whole window
  (91 rows < 100, max row 09:31:47, no paging, the #10 lower-bound caveat not applicable to a
  single-sided query). All rows eyeballed in addition to the keyword filter.
- n8n/SiYuan advisories are repo-level GHSAs: the global-advisory API 404s on them, the repo
  advisory pages render 200 for curl and carry the CVSS vectors; vulncheck advisories + NVD CVE
  pages fetched as corroboration (all 200).
- invisible-dots used for exactly three sources this round: research.jfrog.com/posts list,
  ReversingLabs /blog (its /blog/research-labs path 404s into a "Well, This Is Unexpected"
  shell - the correct entry is /blog + sitemap lastmod), and cursor.com/changelog (per #10,
  www subdomain still times out). The shared main profile was free this time. Per operator
  instruction all invisible-mcp-vnc containers are stopped and removed after the sweep.

## 2026-11-14 00:00 UTC — sweep #12

Window: Nov 13 12:20 UTC -> Nov 14 00:40 UTC (picks up where sweep #11 stopped at 12:20).
Sources swept: OpenAI news (RSS), Anthropic (sitemap), Google Project Zero (feed), Microsoft
Security Blog + Microsoft Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The
Red, Daily Swig, BleepingComputer, arXiv cs.CR+cs.CL (both RSS channels finally non-empty —
lastBuildDate Nov 13 05:00 — plus /list/recent), GitHub global advisories API (591 rows,
cursor-paged to exhaustion; the 12:31 UTC batch is a large NVD import, every row keyword-scanned
and AI-hits eyeballed) + per-repo advisory pages (llama.cpp, vLLM, transformers, LangChain,
CrewAI, PyTorch, NLTK), NVD keyword searches, Hacker News (>=150 pts + keyword scan over all 665
in-window stories), Techmeme (feed.xml), Hugging Face blog, ReversingLabs blog (sitemap lastmod),
JFrog Research (RSS — no browser needed this round, see method notes), GitHub blog + changelog +
Copilot changelog, coding-agent changelogs (Claude Code, Codex, Copilot CLI, Cursor). Every added
item's URL was fetched this session, HTTP 200, product/CVE/GHSA/paper names verified present in
body; every rejected candidate URL was fetched or explicitly dispositioned below.

+added:    Amazon Threat Intelligence — MadPot decoy LLM-inference-API sensors captured ARTEX
  agentic-pentest framework traffic carrying the agents' own campaign memory (Oct 3-Nov 5): a
  self-sustaining AI-model-credential theft operation (stolen keys fund the inference that powers
  further hunting) and a real-time deepfake-injection KYC-bypass campaign against crypto
  exchanges; AWS notified customers and published IOCs — https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/
+added:    IBM ContextForge MCP Gateway — NVD imported eight IBM bulletins (CVE-2026-110178
  cross-tenant overwrite of public-tool endpoint URLs/headers/auth credentials via /v1/import,
  110194 SSRF, 110195 credential-level authorization bypass, 110279 X-Upstream-Authorization
  overrides gateway credentials, 110278 identity-claim replay, 110189 jq-filter-guard denylist +
  94407 ReDoS DoS, 110199 XSS), 1.0.0-1.1.0; #8 already carries the CVE-2026-18484 tool-collision
  advisory from the same product — https://www.ibm.com/support/pages/node/7296701
+added:    Jet Admin Cloud Platform prompt injection (CVE-2026-123965, GHSA critical) — free
  account exposes .env configuration and exfiltrates credentials via prompt injection in all
  versions; mitigated via sandbox ~Oct 29; CISA carries a CSAF note titled 'Jet Admin Cloud
  Platform Prompt Injection' — https://github.com/advisories/GHSA-65p4-8rwx-hgv4
+added:    vLLM six-GHSA batch (CVE-2026-126394..126399, published one NVD batch at 12:31 UTC) —
  request-controlled max_patches/num_frames/min_pixels bypass multimodal preprocessing limits
  (quadratic CPU / OOM the API server), embeddings-only completions trip a fatal assertion on
  XD-RoPE models, and a lowercase-only sanitize_message() regex leaks live heap addresses
  (ASLR-bypass building block) via uppercase PIL addresses in crafted PNGs; five fixed in 0.31.0 — https://github.com/advisories/GHSA-8hqr-48h6-r93j
+added:    Apache APISIX MCP endpoint (CVE-2026-124726) — relative-path traversal lets MCP callers
  reach routes the OpenAPI document never declared, reusing the declared tool's method and the
  Route's credentials; 3.19.0 before 3.20.0, fixed 3.20.0 — https://github.com/advisories/GHSA-rqq9-382j-9fmw
+added:    NLTK batch advisories on the repo advisory list — unbounded/quadratic CPU DoS in
  SteppingRecursiveDescentParser, CCG lexicon parser, ConllCorpusReader, TreebankWordTokenizer,
  LegalitySyllableTokenizer plus save/load_maxent_params path traversal; several carry CVEs
  (126300/126303/126304 etc.), all hitting the global advisory DB on Nov 13 — https://github.com/nltk/nltk/security/advisories
+added:    Project Zero retires 90+30 for 'Full 90' -> 'Full 60' (Feb 2, 2027) -> 'Full 30'
  (Jan 3, 2028), citing LLM-accelerated discovery (several vulns/week where it was a few a year),
  near-zero patch-to-exploitation time, and unmanageable vendor negotiation; grace period and
  variant policy removed (post dated Nov 13 08:00 UTC but absent from the feed at sweep #11's
  fetch — first dispositioned here) — https://projectzero.google/2026/11/disclosure-policy.html
+added:    Claude Code v2.1.325 (Nov 13 20:40 UTC) — 'Fixed a security issue in how
  zstd-compressed responses are decoded', release notes only, no CVE/advisory; same pattern as
  the v2.1.319 (#2) and v2.1.321 (#5) changelog items, distinct from the #10/#11 rejected
  hardening lines because a security issue fix is explicitly disclosed — https://github.com/anthropics/claude-code/releases/tag/v2.1.325
+added:    arXiv 2611.14548 SideTaskBench — third-party technical references steer coding agents
  into producing and DELIVERING independent side-task artifacts during authorized repairs; 25
  repairs x 190 side-task units, four-level delivery scoring (missed by #11's /list/recent scan) — https://arxiv.org/abs/2611.14548
+added:    arXiv 2611.14978 — a disguised LoRA adapter passes audit on the unquantized base yet
  hijacks decision tasks (moderation, function selection) after quantized deployment, no base
  modification, no knowledge of the quantization config (ML supply chain) — https://arxiv.org/abs/2611.14978
+added:    arXiv 2611.16122 Intent-Aware Security — trusted agent intent as a first-class
  security object, catching actions with no attack pattern and within permissions that still
  violate role/organizational purpose — https://arxiv.org/abs/2611.16122
+added:    arXiv 2611.16158 PolicySniff — co-resident GPU tenant infers an agent's confidential
  approval-policy state by timing its own kernels during the agent's processing — https://arxiv.org/abs/2611.16158

-rejected: FAA outage cluster — HN 1,099-pt "FAA issues national ground stop", Techmeme p30-p35
  (fetched p32-p35): Bloomberg-sourced line says the FBI does NOT believe the outage was a
  cyberattack or criminality; CBS-sourced line attributes it to a private-communications-network
  update that rebooted/crashed the ATC system; the only AI angle is ASI's Nov 13 denial that the
  AI tool SMART (whose FAA expansion Techmeme p32 recaps) was involved — no security disclosure,
  availability incident (#10 M365-outage precedent)
-rejected: HN >=150 remainder: $11.4M Nvidia hardware vanishes in shipment (logistics, no cyber),
  Brazil Caixa insurer breach, 7.6M customers (press-release PDF fetched, zero AI/ML factor),
  AI-unlocks-Godel-notebooks (research, not security), Nvidia-petition/Amodei/PowerPoint/Zig/Haskell/
  LibreOffice/manifesto items (corporate/opinion/product); 80-149 keyword band: OpenAI robotics
  disbanding (business), Jonathan Blow video, AI-schools theories — nothing AI-security
-rejected: BleepingComputer in window: Trezor confirms crypto-stealing cellular implants in three
  Safe 5 wallets from one reseller (fetch verified — hardware supply chain, no AI factor),
  German drone-maker Quantum Systems update-server compromise per Spiegel/FT (no AI factor),
  Check Point gateways hacked via PHP2Shell to deploy ESXi wiper (product vuln, no AI factor),
  "How one AI agent exposed five critical enterprise security gaps" (SPONSORED by Token Security —
  marketing content), Microsoft printing issues (not a disclosure)
-rejected: Microsoft Security Blog "SCKit abuses ScreenConnect" (Nov 13 21:31) — thorough worm
  write-up, AI terms appear only in site navigation; no AI factor (precedent: #11 non-AI vulns)
-rejected: OpenAI "Introducing ChatGPT Wallet" (Nov 13 15:00) — product launch, no security
  disclosure (#11 Brave x402 precedent); note: openai.com 403s to the validator UA, dispositioned
  from the RSS entry without body fetch
-rejected: GitHub changelog in window: Manage AI Scan for repositories via REST API (scanning
  product enablement), GitHub Enterprise Server support in the Copilot app, Gemini 4 Hydrogen in
  Copilot, December Copilot model deprecation — product changes, no security fixes (#6-#10 handling)
-rejected: Copilot CLI v1.0.111 (13:40) + v1.0.112 (23:59) full release bodies — feature work
  (regex support in enterprise shell allow-lists) and dozens of UX fixes; the MCP-trust warnings,
  org-policy MCP disable, and terminal-escape-sequence rendering fixes are the same class rejected
  in #4/#5 (no advisory/CVE/security-fix line); Claude Code nothing else in window; Codex
  rust-v0.169.0 stable dispositioned in #11, rust-v0.168.0 (03:54, missed by #11's "first stable"
  claim, body is product features incl. model-provider browsing) and 0.171.0-alpha.1/2/4 bodies
  verified single "Release ..." lines (#6-#11 handling); Cursor changelog now renders for curl —
  newest entry still Nov 6 "Voice Chat with Agents"
-rejected: Anthropic sitemap — site-wide lastmod churn at 17:02 and 22:37 Nov 13 hits ~every URL
  (bulk re-edit, #10 claude-4 lastmod precedent); the one distinct touch is claude-frontier-academy
  — fetched: a Oct 2, 2026 program post ($100M engineer training), and donating-agent-skills is
  Oct 22; neither new, neither a disclosure
-rejected: GitHub global advisories, 591 rows Nov 13 12:20 -> Nov 14 00:05 (cursor-paged; 339 rows
  in the 12:31 NVD-import batch alone) — keyword hits dispositioned above; NLTK rows without CVEs
  are GitHub's own "Duplicate Advisory" markers for the canonical GHSAs and are folded into the
  NLTK item instead of listed separately; non-AI remainder: WordPress-plugin sets, NetScaler
  "Memory overflow" set, Ivanti EPMM, Wasmtime, FrankenPHP, Cudy TR3000 x2, SurveyKing, NocoDB x2,
  Passster, Widget Options, MISP, Rucio, Kroxylicious, SourceCodester set, Linksys, TinyVNC/
  UltraVNC, MCUboot, Cairo-dock, Red Hat Quay, Zipline, plus Hugo Blog AI-assistant-proxy
  unauth-forwarding (CVE-2026-68905, GHSA-pj75-r769-qjv2) rejected as a hobby template repo —
  outside the spec section 2 "major AI repos" advisory scope, impact is free upstream-API abuse —
  and IBM App Connect AI-Agent error-message info leak (CVE-2026-109158, IBM bulletin fetched
  via NVD reference) rejected as product info-disclosure with only naming AI relevance
-rejected: arXiv remainder — the two channels' Friday batch is the same announcement batch sweep
  #11 dispositioned (2611.16693/14864/14808/17115/14970/14950 added, rejections stand); new to the
  RSS and dispositioned now: 2611.14581 evaluator-disagreement study (benchmark methodology, no
  attack), 2611.15302 Browser Immune System (cybersecurity transformer FOR detection =
  AI-for-security, #11 keyword rule), and the replacement stream (2611.04411 TxBench, 2506.15253
  spec-based agent eval, 2606.18532 AI Sandboxes, 2606.20910 agent fingerprinting, 2607.13920
  DeepStress, 2610.39509 jailbreak survey and the rest of the June-October IDs) — age rule,
  re-announcement precedent (#9 Q10); ScrapeBench 2611.11306 legal-compliance not security
-rejected: Meta newsroom / Microsoft Research blog / Unit 42 / Hugging Face blog non-security
  posts: llama.cpp v1.0.0 HF post (ecosystem release milestone, fetched, no security content),
  XYBench + SimReady posts (product/research), everything else on those feeds predates the window

Swept, zero new items in window: OpenAI news (ChatGPT Wallet rejected above, customer stories
pre-window), Meta newsroom (latest Nov 12, dispositioned #10/#11), Microsoft Research blog
(Focus #98 rejected #10), Unit 42 (latest Nov 12, added #10), Embrace The Red (blog index read
directly, latest post still Oct 13), Hugging Face blog incl. malware-scanner posts, GitHub blog
(availability report + agent-skills guide pre-window), per-repo advisory pages llama.cpp / vLLM /
transformers / LangChain / CrewAI / PyTorch (no GHSA newer than those dispositioned in #9-#11; the
vLLM repo page now lists sibling advisory GHSA-m9fg-w9qh-67pc — the #9 "absent from repo list"
note is resolved), ReversingLabs (sitemap lastmod >= Nov 12 only on alliance/glossary/author and
the Nov 4 'Securing Frontier AI' post — re-edits per #11), JFrog Research RSS (newest post Nov 9
Contagious Interview, dispositioned #10/#11; the /post listing itself is JS-only but the RSS at
/rss.xml covers it), NVD keyword searches Nov 13 12:20 -> 00:40: "jailbreak" 0, "prompt injection"
1 (= Jet Admin), "LLM" 1 (= Hugo Blog, rejected), "AI agent" 1 (= App Connect, rejected),
"MCP" 9 (= APISIX + the 8-bulletin ContextForge batch), Daily Swig (domain still does not resolve,
Q1 open).

Method notes (not policy questions):
- arXiv: both RSS channels carried the Friday batch for the first time since #8 (channels had
  been empty five rounds). The batch matches #11's dispositioned set except four security titles
  #11's /list/recent scan missed (SideTaskBench, disguised-adapter hijack, intent-aware
  protection, PolicySniff) — all four were scanned and added here; RSS is now treated as the
  source of truth, /list/recent demoted to fallback.
- GitHub advisories: the published=>=query this window returned 591 rows behind cursor
  pagination (page parameter silently ignored — the endpoint pages via the Link header `after`
  cursor). 339 rows share the 12:31:11-12:31:19 UTC timestamps: GitHub bulk-imported an NVD
  tranche; only keyword hits were escalated, all non-AI families listed in the reject bullet.
- NVD keyword API pubStartDate/pubEndDate windows used instead of #11's heuristic; the 22:17 UTC
  IBM tranche surfaced ContextForge, which the global-GHSA path does not cover (vendor-bulletin
  CVEs without GHSAs) — the two keyword sources are complementary, kept both.
- IBM support pages rate-limited parallel fetches (7 parallel -> all 000); sequential with 4s
  spacing returned 200s. Corroboration for the ContextForge item uses IBM bulletin pages, not an
  aggregator.
- invisible-dots browser was NOT used this round and could not start: browser_open fails, the
  proxy at 127.0.0.1:11080 refuses connections (three egress-check endpoints, instant
  connect_failed). Every source that needed the browser in #10/#11 was read without it:
  cursor.com/changelog renders for plain curl (Nov 6 newest), reversinglabs.com/sitemap.xml
  carries per-post lastmod, and research.jfrog.com/rss.xml covers the post list. No source was
  skipped for lack of a browser. Per operator instruction all invisible-mcp-vnc containers are
  stopped and removed after the sweep.

## 2026-11-14 12:00 UTC — sweep #13

Window: Nov 14 00:40 UTC -> Nov 14 12:00 UTC (picks up where sweep #12 stopped at 00:40).
Sources swept: OpenAI news (RSS), Anthropic (sitemap + both distinct page bodies), Google Project
Zero (feed), Microsoft Security Blog + Microsoft Research blog, Meta newsroom, AWS Security Blog,
Unit 42, Embrace The Red (RSS), Daily Swig, BleepingComputer, arXiv cs.CR+cs.CL (both RSS channels
+ /list/recent fallback), GitHub global advisories API (all 61 in-window rows, keyword re-scan
over FULL descriptions) + per-repo advisory pages (llama.cpp, vLLM, transformers, LangChain,
CrewAI, PyTorch, NLTK, ggml), NVD keyword searches (jailbreak / prompt injection / LLM / "AI
agent" / MCP), Hacker News (all 237 in-window stories: >=150 band + keyword scan), Techmeme
(feed.xml), Hugging Face blog, ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog
+ changelog (full + Copilot label), coding-agent changelogs (Claude Code, Codex, Copilot CLI,
Cursor). Every rejected candidate URL was fetched this session unless its disposition is cited to
an earlier sweep; 0 items added.

-rejected: ggml v1.0.0 (HN 293) — milestone release; full release-notes body fetched (~20 KB),
  zero security-fix/CVE/vulnerability lines (#12 llama.cpp v1.0.0 + HF-post precedent)
-rejected: Citizen Lab "Security Agencies Worldwide Found to Use Ad-Based Surveillance System"
  (HN 113, Nov 12 20:31) — post body fetched: ad-targeting surveillance platform study, zero
  AI/ML factor in the mechanism (no-AI-factor bar)
-rejected: HN >=150 remainder — Nvidia-funding-AI-startups-scarcity claim (656, business), Nvidia
  director $1B stock sale (203, finance), Ukraine-Russia refinery strikes (269, geopolitics),
  Monkey Island versions (187, games), FAA ground-stop lifted (208) + Amtrak booking issues (169)
  — the #12 FAA cluster continuing: availability incidents, no AI security disclosure
-rejected: willdepue X thread "Blackwell signatures in open-weight models" (HN 13) — named
  researcher self-report, but the oEmbed body is truncated at the tweet with no PoC/repo/CVE
  artifact fetchable, thread is Nov 12 and no second source corroborated it in window;
  export-control forensics, not AI-product security (#12 Godel-notebooks precedent)
-rejected: pgaleone.eu "Backdooring LLMs - Do you trust your AI provider?" (HN 2) — the primary
  URL is listed in the site's own index and sitemap yet returns 404 to repeated GET and HEAD
  (cache-busted), so no content could be verified; no corroboration source (spec 3 fetch-200 bar)
-rejected: Google Developers blog "Agent Distraction" (HN 2) — post fetched: token/step/wait-time
  overhead of legacy TODO(b/...) comments across 913k coding-agent sessions; body contains zero
  occurrences of injection/security/malicious/exploit/attack/vulnerability — reliability/efficiency
  analysis, no disclosure (#12 evaluator-disagreement precedent)
-rejected: Bitrise status-page security incident (HN 1) — CI-provider login-service incident,
  fetched incident text: no AI factor (precedent #10 4,800-GitHub-accounts); Airtasker "third
  party security incident" (HN 2) — support page 403s to the validator UA with no second source
  to corroborate (spec 3: 403 without corroboration is dropped)
-rejected: Anthropic sitemap distinct lastmod touches claude-text-watermark (Nov 13 18:29) and
  claude-for-teachers (Nov 13 12:49) — both fetched: datePublished 2026-08-14 and 2026-07-14,
  i.e. re-edits of ~3-month-old posts, not first reports (#10/#12 lastmod-churn precedent); all
  other sitemap touches are the site-wide 17:02/22:37 Nov 13 re-edit churn already dispositioned
  in #12, no URL carries a lastmod after Nov 13 22:37
-rejected: GitHub global advisories Nov 14 00:00-12:00 (61 rows, one page) — the 8 IBM
  ContextForge GHSAs (cvpg-9259-jm3r, v26p-9cf3-34w5, 254c-53h8-vpf6, 5c6v-rx44-457j,
  fp2v-98c9-h9m3, vwg6-jvc2-rpch, r7mv-8vcv-gjv6, 774m-v6g3-r639) published 00:31 UTC are
  GitHub mirrors of the eight IBM bulletins the #12 ContextForge item already carries (same CVE
  families 110178/110194/110195/110279/110278/110189+94407/110199, same 1.0.0-1.1.0 ranges — no
  new facts, item unchanged); GHSA-f65r-qr5q-4ggg IBM App Connect restates the #12 rejection
  (AI-agent error-message info leak, naming-only AI relevance); Devolutions PowerShell "agent
  configuration folder" = software agent, not AI; remainder — WordPress plugin/theme set, abrt
  x2, dbus nonce-tcp, Stirling-PDF, HashiCorp cap JWT lib, Control Web Panel, Postecert FIRMAOK!,
  Johnson Controls default credentials, projectworlds x2, IBM Bob/Cyber-Fraud/DataStage/QRadar/i/
  webMethods naming-zero-AI advisories — no AI factor
-rejected: Techmeme Nov 14 top block — Microsoft 365 all-operational resolution (availability,
  #10/#12 precedent), Jensen Huang/China business story, funding/marketing five (Operand $31M,
  WhiteMatter guardian-agents soft launch, Charta Health, Luminal, Tendrils — spec 3 rejects
  funding/marketing), JD.com/Ceconomy close
-rejected: Copilot CLI v1.0.112-0 prerelease (Nov 13 19:37, missed by #12's body check of v1.0.111
  + v1.0.112) — full body fetched: Ventura startup UX, per-model context-window defaults,
  /chronicle cost tips = the same class rejected in #4/#5/#12, no security-fix line; Codex
  rust-v0.171.0-alpha.5 (01:20) — single "Release ..." line (#6-#12 handling); Claude Code
  nothing since v2.1.325 (added #12)
-rejected: arXiv — both RSS channels empty (weekend gap; the Friday batch was dispositioned in
  #11/#12); /list/recent 50-entry new-submissions list re-scanned per the #12 fallback rule:
  every keyword-title hit was already dispositioned in #11 (2611.16151 fault-injection
  Q-learning, 2611.17280 synthesis smuggling, 2611.17123 + 2611.16565 AI-for-security, 2611.15583
  image-hiding attacks, 2611.15118 Lazare duplicate of the #7 item, rest in the out-of-scope set
  #11 listed) — nothing un-dispositioned remains

Swept, zero new items in window: OpenAI news RSS (latest Nov 13 15:00, dispositioned #12),
Microsoft Security Blog (latest SCKit, rejected #12), Microsoft Research blog (latest Nov 10
Lazare, dispositioned #10/#11), Meta newsroom (latest Nov 12), AWS Security Blog (latest Nov 13
ARTEX post, added #12), Unit 42 (latest Nov 12, added #10), Embrace The Red (RSS latest Oct 13),
Project Zero feed (latest Nov 13 policy post, added #12), BleepingComputer (latest item Nov 13
19:38 UTC, all in-window items dispositioned #11/#12), per-repo advisory pages llama.cpp / vLLM /
transformers / LangChain / CrewAI / PyTorch / ggml (ggml has zero advisories; no repo advisory
newer than those dispositioned in #9-#12), NVD keyword searches Nov 14 00:40 -> 12:00: jailbreak
0, "prompt injection" 0, LLM 0, "AI agent" 0, MCP 0, Hugging Face blog (Nov 13 posts dispositioned
#12), JFrog Research RSS (newest Nov 9), ReversingLabs sitemap (only alliance/solutions/author
page re-edits, no post), GitHub blog + changelog (nothing newer than the Nov 13 entries
dispositioned #12), Cursor changelog (newest entry still Nov 6), Daily Swig (domain still does not
resolve, Q1 open).

Method notes (not policy questions):
- arXiv: RSS stayed empty (no announcement batch between the Friday one #12 consumed and Monday's);
  the /list/recent fallback confirmed no missed title — the keyword-hit set is exactly #11's
  dispositioned set, so this round needed no abs-page additions.
- GitHub advisories: keyword scan this round ran over each advisory's FULL description, not just
  the summary + 600 chars used in #12 — same 10 escalated rows, so the #12 method has no misses on
  this tranche; the 00:31 UTC IBM ContextForge GHSA batch shows the NVD-to-GitHub advisory mirror
  lag is now ~2h for IBM bulletins, so next rounds should keep both keyword paths (GitHub + NVD).
- HN window: created_at_i 1794616800-1794657600 (Nov 14 00:40-12:00 UTC), 237/237 stories scanned;
  note the >=150 band is thin on Saturday — the keyword band caught the only real candidates
  (Citizen Lab, pgaleone, Bitrise/Airtasker), all dispositioned above.
- invisible-dots browser was NOT needed: every source in this window (Anthropic news HTML,
  cursor.com changelog, reversinglabs sitemap, research.jfrog RSS, publish.twitter.com oEmbed for
  the Blackwell thread) answered plain curl. Per operator instruction all invisible-mcp-vnc
  containers are stopped and removed after the sweep (none were running at start or end).

## 2026-11-15 00:20 UTC — sweep #14

Window: Nov 14 12:00 UTC -> Nov 15 00:20 UTC (picks up where sweep #13 stopped at 12:00).
Sources swept: OpenAI news (RSS), Anthropic (sitemap), Google Project Zero (feed.xml), Microsoft
Security Blog + Microsoft Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red
(blog index.xml), Daily Swig, BleepingComputer, arXiv cs.CR+cs.CL (both RSS channels +
/list/recent?show=250 fallback), GitHub global advisories API (all 40 in-window rows, keyword scan
over summary + full description) + per-repo advisory pages (llama.cpp, vLLM, transformers,
LangChain, CrewAI, PyTorch, ggml), NVD keyword searches (jailbreak / prompt injection / LLM /
"AI agent" / MCP), Hacker News (all 435 in-window stories: >=150 band + keyword scan), Techmeme
(feed.xml), Hugging Face blog, ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog
+ changelog, coding-agent changelogs (Claude Code, Codex, Copilot CLI, Cursor). 1 item added,
13 disposition lines below; every candidate URL was fetched this session unless dispositioned by
an earlier sweep.

+added:    Cursor incident update — Fireworks-exposed access key, agent-trace credentials
  (<0.01% of customers), archived source repos and support data reached, malicious Claude Code
  extension copy closed out; vendor primary post Nov 13 12:00 UTC, BleepingComputer corroboration
  (its "up to 5,000 developers" backdoored-extension estimate noted as the reporter's figure) —
  https://cursor.com/blog/security-incident-update (HEAD + GET 200; Cursor, Claude Code extension,
  Fireworks all present in body; missed by #12 only because that round swept the Cursor changelog,
  not cursor.com/blog — blog is now on the per-round list)
-rejected: crissyfield.de "Lovable: Many Apps Built on the Platform Expose Sensitive Data"
  (HN 586, Nov 14 front page) — updated English report from a named researcher (passport data,
  200K+ messages), but page Last-Modified is Nov 3 and the researchers' own Handelsblatt coverage
  is Nov 2: the disclosure is 11-12 days old and was not first reported in this window, so spec 3
  age bar fails and the >72h exception does not apply; Q11 (questions.md) still governs whether
  such pre-machine in-scope findings may ever be backfilled — top candidate if Q11 is answered yes
-rejected: US/DPRK "FaceStealer" disruption (BleepingComputer Nov 14 17:27 UTC) — 18-month crypto
  theft campaign taken down; mechanism is social engineering with deepfaked video meetings,
  ClickFix and malware, targets are crypto/Web3 users; zero occurrences of AI in the article body
  as a system involved (no-AI-factor bar, #13 Citizen Lab precedent)
-rejected: EFF "Tech companies are trying to shift the concept of privacy again" (HN 435) —
  advocacy essay on privacy framing around smart glasses / on-device AI features; no disclosure,
  no PoC/CVE/repo; spec 3 opinion bar requires a named researcher plus evidence artifact
-rejected: Garante EUR 805M fine on Meta over Android users' data access (Techmeme + HN 307) —
  data-privacy enforcement decision, no AI-product security factor (#10/#12 precedent class)
-rejected: PyPI "redesigned project pages with clearer security signals now live" (HN 193) —
  blog.pypi.org answered HTTP 429 to all three fetch attempts (Chrome UA and plain curl UA), so
  the primary page could not be fetched this session (spec 5 fetch-200 bar); also the underlying
  post is dated Nov 6 (age >72h, not a first report) and a security-signal UI launch is the
  feature-announcement class already rejected in #12 ("Manage AI Scan REST API")
-rejected: GHSA-c2gc-chxg-hvqr / CVE-2026-129428 Pulse "missing authorization in
  /api/ai/approvals" (published Nov 14 15:31 UTC) — rcourtman/Pulse is a Proxmox VE / PBS / Docker
  monitoring dashboard whose README documents no AI assistant; the AI relevance lives only in the
  endpoint path, mechanism is a generic horizontal-privilege flaw between authenticated users
  (#12 IBM App Connect naming-only precedent)
-rejected: GitHub global advisories Nov 14 12:00 -> Nov 15 00:00 (40 rows, one page, full
  descriptions scanned) — FRRouting bgpd, Perl (Dancer, HTTP::Session, JSON::XS x7), WordPress
  plugin set (FooBox, BuddyForms, Contact Form 7, Shipo, MemberDash, Cloud Storage Manager),
  Kingston FURY CTRL, CoolerMaster MASTERPLUS+, Tenda, Edimax, Ettercap x2, Dokploy, Kroki,
  sox_ng, XBackBone, Wubang, VanBlog x2, Axeloz, Frappe, angular-editor x2, Akaunting, RuoYi,
  CloudSchool, Opixido OCMS, grayback ATO-XSS advisory (GHSA-ffwr-vmf2-gpp2) — zero AI/ML factor
  in every row besides the Pulse row above
-rejected: Copilot CLI v1.0.112 stable (Nov 13 23:59 UTC, fell between the #12 body check and
  #13's v1.0.112-0 prerelease check; body re-fetched this round) — sandbox host-rules help text and
  emoji/CJK path truncation fixes, zero security-fix lines (#4/#5/#12/#13 bar); Codex
  rust-v0.171.0-alpha.5 dispositioned in #13; Claude Code nothing since v2.1.325 (added #12)
-rejected: HN >=150 remainder — SpaceX Starship Flight 15 catch (515), KDE documentary (360),
  software-job-title family tree (183), LinkedIn/Monster MLM fake-job takedown after CBC
  investigation (164, consumer job-listing fraud, no AI factor), Google Antigravity /teamwork
  research post (154, multi-agent coding feature announcement, zero security disclosure)
-rejected: Techmeme Nov 14 — Jensen Huang resignation cluster (petition past 100K, investor
  letter, China/open-model fallout, GPU-smuggling angle), Italy Meta fine (above), USD-stablecoin
  LatAm share, Databento funding, Trump Media bond deadline, Tesla Cybercab subscription terms,
  FAA outage cause statement (availability, #12 precedent) — business/finance/availability, spec 3
-rejected: HN keyword-band remainder — Show HN agent/tool launches (Agweb, Shotoku, Cyberviber,
  Bright Lines, Ryph, Timon, Cowork, Mopup, CodeRelay), essays (ring-0 protections, analysis by
  synthesis), Medium prompt-injection musing, agent-permission notes — zero vulnerability
  disclosures with artifacts; spec 3 opinion bar
-rejected: arXiv undispositioned keyword-title hits from the Friday Nov 13 batch, re-scanned via
  /list/recent?show=250 (both RSS channels empty, weekend gap) — 2611.17065 FPGA anti-tamper
  sensor security (hardware, no AI factor), 2611.16890 multi-bit LLM watermarking (provenance,
  outside spec 2 keywords), 2611.15013 LLM-unlearning privacy defense (unlearning), 2611.14984
  VRMaaS verifiable reward models (alignment infra), 2611.15102 fake-news input shifting LLM
  forecasts (misinformation study); plus 2611.14769 "Leaderboard Hacking" (HN 1): primary category
  cs.AI, outside spec 2 cs.CR+cs.CL scope, submitted Nov 10 (age >72h) and a 1-point HN link is
  not first-report coverage

Swept, zero new items in window: OpenAI news RSS (latest Nov 13 15:00, dispositioned #12),
Anthropic sitemap (no lastmod after Nov 13 22:37, churn dispositioned #12/#13), Project Zero
feed.xml (latest Nov 13 policy post, added #12; old /feeds/posts/default path is 404 — feed.xml is
the live one, found on the site), Microsoft Security Blog (SCKit, rejected #12), Microsoft Research
blog (latest Nov 10), Meta newsroom (latest Nov 12), AWS Security Blog (Nov 13 ARTEX, added #12),
Unit 42 (latest Nov 12), Embrace The Red (index.xml latest Oct 13; /blog/rss.xml path now 404),
BleepingComputer (two in-window items dispositioned above, rest pre-window), per-repo advisory
pages llama.cpp (newest Mar 2026) / vLLM (Nov 2) / transformers (none listed) / LangChain (none
newer) / CrewAI (none) / PyTorch (Nov 10) / ggml (zero) — nothing newer than #9-#13 dispositions,
NVD keyword searches Nov 14 12:00 -> Nov 15 00:00: jailbreak 0, "prompt injection" 0, LLM 0,
"AI agent" 0, MCP 0, Hugging Face blog (nothing after Nov 13), JFrog Research RSS (newest Nov 9),
ReversingLabs sitemap (only a Nov 14 01:21 alliance-page re-edit, inside #13's window and
dispositioned there; nothing after 12:00), GitHub blog + changelog (Oct availability report =
availability-only post, Copilot model deprecation + GHES/Gemini model additions = lifecycle and
features, no security disclosure), Cursor changelog (newest entry still Nov 6), Daily Swig (domain
still does not resolve, Q1 open).

Method notes (not policy questions):
- HN window: created_at_i 1794657600-1794698400 (Nov 14 12:00 -> Nov 15 00:00 UTC), 435/435
  stories scanned; Sunday bands are thin, the two real candidates (Lovable, Cursor) both came via
  the >=150/keyword bands and are dispositioned above.
- arXiv: Sunday means no announcement batches between the Friday one #11/#12 consumed and
  Monday's; the show=250 list fallback confirmed every Friday-batch keyword hit is now
  dispositioned (the five undispositioned ones are rejected above line-by-line).
- GitHub advisories: this window's 40 rows fit one page; both summary and full description were
  keyword-scanned per the #13 method note, so no description-only misses in this tranche.
- invisible-dots browser was NOT needed: every source in this window answered plain curl (the one
  429 source, blog.pypi.org, also fails the age bar, so a browser attempt would not change the
  disposition). Per operator instruction all invisible-mcp-vnc containers are stopped and removed
  after the sweep (none were running at start or end).

## 2026-11-15 12:12 UTC — sweep #15

Window: Nov 15 00:00 UTC -> Nov 15 12:00 UTC (picks up where sweep #14 stopped at 00:20).
Sources swept: OpenAI news (RSS), Anthropic (sitemap), Google Project Zero (feed.xml), Microsoft
Security Blog + Microsoft Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red
(blog index.xml), Daily Swig (reachability retest), BleepingComputer (main feed; the
news/artificial-intelligence/feed path now answers 404), arXiv cs.CR+cs.CL (RSS channels +
/list/<cat>/recent fallback), GitHub global advisories API (newest-first sweep of ~230 rows across
pages, summary + description keyword scan) + per-repo advisory listings via API (llama.cpp, vLLM,
transformers, LangChain, CrewAI, PyTorch, ggml), NVD keyword searches (jailbreak / prompt injection
/ LLM / "AI agent" / MCP), Hacker News (all 261 in-window stories: >=150 band + keyword scan),
Techmeme (feed.xml, 10 in-window items), Hugging Face blog, ReversingLabs (sitemap lastmod), JFrog
Research (RSS), GitHub blog + changelog, coding-agent changelogs (Claude Code, Codex, Copilot CLI,
Cursor). 1 item added, 9 disposition lines below; every candidate URL was fetched this session
unless dispositioned by an earlier sweep.

+added:    FBI FLASH-20261113-001 "Threat Overview of ARTEX Autonomous Penetration Testing
   Framework" (TLP:CLEAR, dated Nov 13) — first government-level consolidation of the ARTEX
   incident cluster: weaponization since at least August 2026 across ten sectors; some intrusions
   ran from initial targeting to privileged access in under three hours; South Korean financial
   campaign exfiltrating personal information of ~67,000 individuals; ConnectWise Cloud Backup
   supply-chain compromise reaching more than 2,250 downstream tenants; >1,600 publicly exposed
   ARTEX consoles in early November; repo taken down by developer "Autumn-27" Oct 8 but forks,
   mirrors and commercial ARTEX PRO keep it available; MITRE-mapped TTP table plus IP/domain/
   behavioral IOCs (LLM-API egress signatures) with CrowdStrike/Microsoft/Censys/Zscaler/Check
   Point/GTIG/Akamai sourcing — https://www.ic3.gov/CSA/2026/261113.pdf (HEAD + GET 200 this
   session, body fetched via pdftotext; FBI/FLASH/ARTEX/"under three hours" all present; first
   surfaced inside my sources via a 3-point HN post at Nov 15 00:47, after #14's 00:20 cutoff;
   government advisory is itself the primary source, no corroboration attached)
-rejected: "Lazarus: Backdooring LLMs by editing ten neurons" (HN 1 pt, Nov 15 00:47 window post)
   — tweet fetched via the syndication API: authored Nov 10 19:59 UTC by Dimitris Papailiopoulos,
   linking an X article previewing the dormant-neuron weight-reprogramming study already carried
   from Microsoft's Security Blog in #7 (research disclosed Nov 10, age >72h, not first reported
   this window); #9/#11 duplicate-rejection precedent applies
-rejected: "Skyfall Attack: Microarchitectural Data Sampling on AMD" (HN 1 pt) — project page
   fetched: transient-execution side channel on AMD's cache way predictor (SEV-SNP guest-state
   extraction, microcode/SMM leaks, CVE-2026-40671, IEEE S&P 2027 paper); solid disclosure but
   zero AI/ML factor in the mechanism or targets (no-AI-factor bar, #13/#14 precedent)
-rejected: TechCrunch "Where was Nvidia's security team?" (HN 3 pts) — URL fetched returns TechCrunch's
   404 template; commentary on the Nvidia/Jensen Huang export-control scandal, no named-researcher
   disclosure or evidence artifact; spec 3 opinion bar
-rejected: EFF "We're Asking an Appeals Court to Require a Warrant for ALPR Searches" (HN 247, top
   >=150 band) — litigation-advocacy post on surveillance policy; no AI-product security
   disclosure, no PoC/CVE/repo; spec 3 opinion bar
-rejected: "Court denies motion to dismiss DRAM price fixing case" (HN 215, >=150 band) — antitrust
   litigation news hosted on litnews.ai; the only AI element is the hosting domain; zero security
   content
-rejected: Techmeme in-window (10 items) — Jensen Huang resignation/scandal cluster (WSJ SE-asia
   chip-access business piece, insider-behavior reports, White House political angle, Wojcicki
   comment), first AI Music Awards production coverage, Nidus defense-robotics raise, TSMC $500M
   fund, Nebuly agent-analytics raise, Applied Digital tenant-guarantee report — business/finance/
   culture, spec 3; none is a security disclosure
-rejected: HN keyword-band remainder — LeakRadar (ad-pixel scanner product), Turbo-Browser-MCP
   (Show HN, 38 pts, MCP client tool launch), Embiody BrowserDo (browser-for-agents preview),
   Nano-Edge-TTS / Vantage-V-AI / quills (Show HN launches), "A Pile of MCP Connectors Is Not a
   Context Engine" (architecture essay), "AI Slop Papers Published Under Prominent Academics
   Names" (publishing-misconduct story, no security artifact), Ask HN node.js security layer
   (question thread) — zero vulnerability disclosures with artifacts; spec 3 opinion bar
-rejected: GitHub global advisories, Nov 15 00:00 -> 12:00 window — zero advisories published in
   window (verified by newest-first sweep: global newest reviewed row is the NLTK DoS batch at
   Nov 14 00:08:50Z, i.e. nothing newer exists; weekend quiet). Keyword hits in the newest ~230
   rows all dispositioned earlier or non-AI: russh agent-lock advisory + Loofah sanitizer set
   (generic software), OpenTelemetry traceparent rows ("MCP" regex false positive on
   "traceparent"), pgEdge RAG/LLM rows (added/rejected in #11), Microsoft Visual Studio/.NET/
   Aspire rows (naming-only, no AI factor)
-rejected: coding-agent changelogs — Copilot CLI (newest v1.0.112 Nov 13 23:59, dispositioned
   #14), Claude Code (newest v2.1.325 Nov 13, added #12), Codex (newest rust-v0.171.0-alpha.5
   Nov 14 01:20, dispositioned #13), Cursor changelog (newest entry still Nov 6): zero new
   releases or security-fix lines in window

Swept, zero new items in window: OpenAI news RSS (latest Nov 13 15:00, dispositioned #12),
Anthropic sitemap (no lastmod on Nov 14-15), Project Zero feed.xml (latest Nov 13, dispositioned
#12), Microsoft Security Blog (latest Nov 13 SCKit, rejected #12), Microsoft Research blog (latest
Nov 10), Meta newsroom (latest Nov 12), AWS Security Blog (latest Nov 13 MadPot, added #12),
Unit 42 (latest Nov 12), Embrace The Red (index.xml latest Oct 13), Hugging Face blog (latest
Nov 13 18:00, llama.cpp v1.0.0 release post is not a security disclosure), JFrog Research RSS
(latest Nov 9), ReversingLabs sitemap (no lastmod on Nov 15; latest Nov 14 01:21, dispositioned
#13/#14), GitHub blog + changelog (changelog latest Nov 13 20:40, dispositioned #14), BleepingComputer
(main feed answered 200, zero posts since Nov 15 00:00), per-repo advisory listings llama.cpp
(newest Mar 2026) / vLLM (Nov 2) / PyTorch (Nov 10, dispositioned #12) / LangChain (Jun 12) /
transformers, CrewAI, ggml (zero listed) — nothing newer than #9-#14 dispositions, NVD keyword
searches Nov 15 00:00 -> 12:00: jailbreak 0, "prompt injection" 0 (first attempt 503, retried
clean), LLM 0, "AI agent" 0, MCP 0, arXiv: no announcement batch in window (both RSS channels
empty; /list/<cat>/recent latest listing is still Fri Nov 13, fully dispositioned #11/#12/#14;
Sunday gap, next batch Monday).
Daily Swig status change: dailyswig.com now resolves (to portswigger.com's IPs) but the TLS
certificate is for another hostname, so the feed is still unfetchable; portswigger.net/dailyswig
redirects to the PortSwigger root and portswigger.net/dailyswig/rss answers 404 — no live Daily
Swig feed reachable, Q1 stays open with updated facts.

Method notes (not policy questions):
- HN window: created_at_i 1794700800-1794744000 (Nov 15 00:00 -> 12:00 UTC), 261/261 stories
  scanned; the >=150 band held exactly 2 stories, both rejected above.
- GitHub advisories API: published-range filters are ignored on /advisories (a query bounded to
  Nov 15 returned Nov 11-14 rows), so window-emptiness was proven instead by a newest-first
  descending sweep; first-page top published_at is the global maximum.
- The FBI FLASH URL answers HEAD 200 + GET 200 to plain curl from this datacenter IP; no
  invisible-dots browser session was needed this round. Per operator instruction all
  invisible-mcp-vnc containers are stopped and removed after the sweep (none were running at
  start or end).

## 2026-11-16 00:22 UTC — sweep #16

Window: Nov 15 12:00 UTC -> Nov 16 00:00 UTC (picks up where sweep #15 stopped at 12:00).
Sources swept: OpenAI news (RSS), Anthropic (sitemap), Google Project Zero (feed.xml), Microsoft
Security Blog + Microsoft Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red
(blog index.xml), Daily Swig (reachability retest), BleepingComputer, arXiv cs.CR+cs.CL (both RSS
channels + /list/<cat>/recent fallback), GitHub global advisories API (newest-first sweep + keyword
scan of summaries and descriptions) + per-repo security-advisories API (llama.cpp, vLLM,
transformers, LangChain, CrewAI, PyTorch, ggml), NVD keyword searches (jailbreak / prompt injection
/ LLM / "AI agent" / MCP), Hacker News (all 538 in-window stories: >=150 band + keyword scan incl. a
sub-2-point security-keyword pass), Techmeme (feed.xml), Hugging Face blog, ReversingLabs (sitemap
lastmod), JFrog Research (RSS), GitHub blog + changelog, coding-agent changelogs (Claude Code,
Codex, Copilot CLI, Cursor). 1 item added, 13 disposition lines below; every candidate URL was
fetched this session unless dispositioned by an earlier sweep.

+added:    CVE-2026-125771 privilege escalation in "MCP Server for WordPress" (AtlasAiDev), the
   plugin that exposes a WordPress site as a Model Context Protocol server for AI agents —
   Patchstack advisory (reported Nov 10 by "iml1s", published Nov 13) covers versions through 2.0.3,
   fixed in 2.0.4; added to NVD in-window Nov 15 15:16 UTC; GitHub advisory mirror attached as
   corroboration — https://patchstack.com/database/wordpress/plugin/ai-workflow-automation-ai-agent-hub/vulnerability/wordpress-mcp-server-for-wordpress-plugin-2-0-3-privilege-escalation-vulnerability
   (HEAD + GET 200 this session; CVE ID + product name present in body; the MCP/agent-connector
   product is the flaw surface itself, not naming-only — Pulse/Aspire precedent distinguished;
   Nov 13 advisory is inside the 72h bar)
-rejected: Claroc Research "GhostShares 2: Grok Apps Exposed xAI's Business Secrets and People's
   Private Data" (HN 5 pts, in-window post Nov 15 17:30) — in-scope disclosure by a named researcher
   (Asaf Aprozper: 43,551 grok.me addresses tested, 35,578 open with no sign-in, incl. 24 SpaceXSI/X
   staff apps; The Hacker News and Cloud Security Alliance coverage located this session), but the
   research published Nov 10: age >72h and the spec 3 exception does not apply because it was not
   first reported in this window — the same URL first surfaced in our HN stream on Nov 10 21:32 UTC
   (6 pts, inside sweep #6's window, under the >=150-only band #6 ran at that time). bREACHYing (#10)
   and Lovable (#14) precedents applied; governing policy question Q11 in questions.md is
   still unanswered, so the age bar stands unchanged. Top backfill candidate if Q11 is answered yes.
-rejected: "Uber confirms hackers stole drivers' license information in breach" (HN 36; TechCrunch
   URL answers its 404 template, story verified via The Hill/Reuters copies) — third-party vendor
   breach at AI-infrastructure provider Fireworks.ai: stolen vendor credentials exposed ride-hailing
   driver license data; the AI vendor is the breached location, not the attacked system (zero AI
   factor in mechanism, targets or attacker tooling; #1 "AI brands as bait" + #14 TSMC class); the
   Fireworks intrusion is already carried in the feed via the Cursor incident update (#14);
   underlying state breach notices trace to the Oct 9 vendor notification
-rejected: "All AI-generated code should be Open Source" (HN 1053, top of >=150 band) —
   policy/opinion essay; spec 3 opinion bar requires a named researcher plus PoC/repo/CVE evidence
-rejected: int8.io "Anatomy of a trojan: backdooring an LLM by editing few dormant neurons" (HN 1
   pt, caught by the sub-2-pt keyword pass) — third-party technical commentary on the
   dormant-neuron backdoor study already carried from Microsoft's Security Blog in #7;
   #9/#11/#15 duplicate-rejection precedent
-rejected: Zenodo "Prompt Injection Attacks in Large Language Model Systems" (HN 3 pts) —
   self-archived Zenodo preprint (record 21357273), no venue, no PoC/repo/CVE artifact; spec 3
   evidence bar; outside spec 2 paper scope (arXiv cs.CR+cs.CL)
-rejected: BleepingComputer "FBI warns ARTEX AI hacking tool is powering global cyberattacks"
   (in-window Nov 15 18:29 UTC) — coverage of FBI FLASH-20261113-001, already added in #15 from the
   primary PDF; duplicate
-rejected: BleepingComputer "TSMC employee data breach linked to cloud provider" (in-window Nov 15
   15:23 UTC) — third-party cloud-provider data breach, zero AI factor
-rejected: "Microsoft 365 outage leads to chlorine gas poisoning of 15 hotel guests" (HN 337) —
   article body fetched (WDR): Outlook/M365 outage delayed a pool-plant alert; availability
   incident, zero AI mentions, no AI-product security disclosure; #12 availability precedent
-rejected: OpenAI in-window posts "A step toward defining the Standard Model on a lattice" and "Our
   new blog for science" (both Nov 15 22:00 UTC) — physics research + blog launch, not security
-rejected: HN >=150 remainder — MacBook Pro OLED PWM (397), positional-embeddings explainer (191),
   Turbo Pascal sizes (187), Emacs-as-browser (169), El Nino explainer (152) — outside niche
-rejected: Techmeme in-window (Nov 15 afternoon, 15 items) — Anthropic IPO ~7x oversubscribed,
   Vulcan/Leo launch failure + follow-ups, Huawei Horinger analysis, Jensen Huang comeback talks,
   Nvidia smuggling charge/appeal, Bessent letter, Pax Silica chip-export list, Lutnick items, Apple
   smart-glasses plan, AI Music Awards, post-Huang Nvidia analysis, Nidus raise — business/politics/
   culture, spec 3; none is a security disclosure
-rejected: coding-agent changelogs — Codex rust-v0.171.0-alpha.6 (Nov 15 18:58) and alpha.7
   (Nov 15 23:13) fell in window, release bodies empty ("Release 0.171.0-alpha.N"), zero
   security-fix lines; Claude Code nothing since v2.1.325 (#12), Copilot CLI nothing since v1.0.112
   (#14), Cursor changelog newest entry still Nov 6

Swept, zero new items in window: OpenAI news RSS (two in-window posts dispositioned above),
Anthropic sitemap (newest /news lastmod still Nov 13 22:37, churn dispositioned #12/#13), Project
Zero feed.xml (latest Nov 13, dispositioned #12), Microsoft Security Blog (latest Nov 13 SCKit,
rejected #12), Microsoft Research blog (latest Nov 12 newsletter edition, not security), Meta
newsroom (latest Nov 12), AWS Security Blog (latest Nov 13 ARTEX, added #12), Unit 42 (latest Nov
12), Embrace The Red (index.xml latest Oct 13), Hugging Face blog (latest Nov 13 18:00,
dispositioned #15), JFrog Research RSS (latest Nov 9, dispositioned #10/#15), ReversingLabs sitemap
(no re-edit after Nov 14 01:21, dispositioned #13/#14), GitHub blog + changelog (latest Nov 13
20:40, dispositioned #14), per-repo advisory listings: llama.cpp (newest Mar 2026) / vLLM (Nov 2,
#1) / transformers (none listed) / LangChain (Jun 12) / CrewAI (zero) / PyTorch (Nov 10,
added #6) / ggml (zero) — nothing newer than #9-#15 dispositions, GitHub global advisories: zero
advisories published Nov 15 12:00 -> Nov 16 00:00 (newest-first sweep; global newest reviewed row
is still the NLTK batch at Nov 14 00:08:50Z; keyword hits in the newest 100 all dispositioned
earlier or non-AI, same list as #15), NVD keyword searches Nov 15 12:00 -> Nov 16 00:00: jailbreak
0, "prompt injection" 0 (first attempt 503, retried clean), LLM 0, "AI agent" 0, MCP 1 (the
WordPress MCP Server row, added above). Daily Swig unchanged: dailyswig.com feed still fails the
TLS handshake from this host, Q1 stays open. arXiv: the Monday announcement batch had not been
published at sweep time (both RSS channels empty; /list/<cat>/recent still shows the Fri Nov 13
listing, fully dispositioned #11/#12/#14; the Monday batch publishes Nov 16 01:00 UTC, one hour
after this window's cutoff — first job of the next round).

Method notes (not policy questions):
- HN window: created_at_i 1794744000-1794787200 (Nov 15 12:00 -> Nov 16 00:00 UTC), 538/538 stories
  scanned; the >=150 band held 7 stories; 163 sub-2-point stories got an extra security-keyword
  pass so a named-researcher self-publication could not hide at 1 point (1 hit, rejected above).
- The GhostShares 2 age finding rests on an Algolia history search over the URL itself: first HN
  appearance Nov 10 21:32:58Z, inside sweep #6's window — a resurfacing at 5 pts five days later
  does not make it a first report.
- GitHub per-repo advisory listings: the working unauthenticated endpoint is
  /repos/{owner}/{repo}/security-advisories (the /repos/.../advisories path answers 404).
- NVD window queries use pubStartDate/pubEndDate with an explicit +00:00 offset; 3 of 5 first
  attempts answered 503 and were each retried to a clean result.
- The invisible-dots browser was NOT needed: every source answered plain curl this round. Per
  operator instruction all invisible-mcp-vnc containers are stopped and removed after the sweep
  (none were running at start or end).

## 2026-11-16 12:20 UTC — sweep #17

Window: Nov 16 00:00 UTC -> Nov 16 12:00 UTC. Sources swept: OpenAI news (RSS), Anthropic
(sitemap), Google Project Zero (feed.xml), Microsoft Security Blog + Microsoft Research blog,
Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red (index.xml), Daily Swig (reachability
retest), BleepingComputer (main feed), arXiv cs.CR+cs.CL (both RSS channels still empty +
/list/<cat>/recent incl. Monday announcement batch via page 2), GitHub global advisories API
(newest-first sweep, 300 rows fetched, keyword scan of summaries + descriptions) + per-repo
security-advisories API (llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch, ggml), NVD
keyword searches (jailbreak / prompt injection / LLM / "AI agent" / MCP), Hacker News (all 333
in-window stories: >=150 band + keyword scan incl. a sub-2-point security-keyword pass), Techmeme
(feed.xml, 15 in-window items), Hugging Face blog, ReversingLabs (sitemap lastmod), JFrog Research
(RSS), GitHub blog + changelog, coding-agent changelogs (Claude Code, Codex, Copilot CLI, Cursor).
7 items added, 20 disposition lines below; every candidate URL was fetched this session unless
dispositioned by an earlier sweep. The Monday arXiv batch (sweep #16 hand-off) is today's payload:
6 papers added from it.

+added:    Gcore Parallax "We found 1,698 Internet-visible AI offensive-security endpoints" —
   Nov 3 measurement of exposed AI offensive-security consoles: 1,255 ARTEX (up from three
   matching services in June 2026), 403 PentAGI, 40 Cairn, across 1,590 IPs; 101 PentAGI endpoints
   share one TLS certificate + frontend (common deployment); cross-referencing the inventory with
   Gcore WAF telemetry surfaced suspicious requests reaching applications in several industries —
   https://gcore.com/blog/ai-offensive-security-endpoints-artex-pentagi-cairn (GET 200 this
   session; 1,698/ARTEX/PentAGI/Cairn/Gcore Parallax all present in body). Research dated
   Nov 9: age >72h, so admitted under the spec 3 exception — first report happened in this window
   (own HN post Nov 16 09:26 UTC is the URL's first-ever HN appearance per an Algolia history
   search; a web search for earlier coverage of the study surfaced nothing) and the vendor's own
   threat-intel post is the primary source
+added:    arXiv 2611.17818 Detail Guidance Attack — jailbreak requests written to look like
   normal user queries pass LLM filters; aux-LLM-composed single-turn requests hit >95% ASR in
   several of 11 victim models x 3 datasets (submitted 11-13, announced in the Monday batch =
   first publication, the paper is the primary source) — https://arxiv.org/abs/2611.17818
+added:    arXiv 2611.17395 "Can LLMs Write Backdoors?" — general-purpose LLMs as weight-space
   editors inject backdoors into target nets from natural-language instructions with no task data,
   forward passes or gradients; 5 LLMs x 3 architectures (MLP/MNIST, ResNet-18/CIFAR-10, Qwen)
   (submitted 11-12, announced in the Monday batch, first-publication exception) —
   https://arxiv.org/abs/2611.17395
+added:    arXiv 2611.17883 Securing Multi-Agent Systems by Design — ReMA optimized malicious
   subtask responses steer the aggregate of a multi-agent system (targeted + untargeted), SecMAS
   defends with semantic-similarity aggregation (announced in the Monday batch) —
   https://arxiv.org/abs/2611.17883
+added:    arXiv 2611.18133 LinEra — agent memory keeps summaries/facts/embeddings the model
   wrote, so message deletion leaves derivatives behind; puncturable-PRF key lineage makes one
   puncture render every descendant unreadable even after full storage + current-key compromise
   (announced in the Monday batch) — https://arxiv.org/abs/2611.18133
+added:    arXiv 2611.17846 "Is Abliteration Getting Harder? Not Uniformly" — Heretic 2.0
   direction-based abliteration + ARA still strip refusal on Qwen3.8-27B, Gemma 4, Olmo 3.1 and
   Mistral Small 3.2 release pairs; WildGuard judged 972 held-out harmful prompts
   (announced in the Monday batch) — https://arxiv.org/abs/2611.17846
+added:    arXiv 2611.18104 Contagious fingerprints — a model continuing another model's
   conversation inherits its text fingerprint; substitute detection drops 0.97->0.72 (DeepSeek for
   Claude Sonnet) and 0.87->0.54 (same-vendor Haiku downgrade), retraining fixes only seen
   substitutes — silent relay handoffs defeat the model-substitution audits they rely on
   (announced in the Monday batch) — https://arxiv.org/abs/2611.18104

-rejected: OpenAI in-window posts "LVMH's Media Division announces strategic partnership with
   OpenAI" (Nov 16 09:00) and "The King's Trust and OpenAI launch £1m partnership" (Nov 16 00:00)
   — business/charity announcements, zero security content
-rejected: Meta newsroom "94% of Singapore Small Business Owners Surveyed Say More AI Guidance Is
   Needed" (Nov 16 06:25) — survey/marketing post, not a security disclosure
-rejected: BleepingComputer "Anthropic's Claude Fable 5.2 blocks fewer security tasks, while
   Mythos gets better at finding bugs" (in-window Nov 16 11:00) — angle on the Fable/Mythos 5.2
   launch; the cited primary page /claude-fable-and-mythos-5-2 is dated November 9 and was
   dispositioned in #7, the sitemap shows no Anthropic re-edit after Nov 13 22:37 — age >72h, not
   first reported in this window (GhostShares #16 age precedent), and a product/safeguards rollout
   is not a vulnerability disclosure (#11 product-page precedent)
-rejected: "Australian retail giant Coles investigates cybersecurity incident" (BC in-window
   Nov 16 11:18 + HN 110 pts + colesgroup.com.au update fetched) — retail incident, zero AI factor
   in mechanism or targets (#1 "AI brands as bait" / #16 TSMC class)
-rejected: BleepingComputer "New NoisyTCP vulnerability lets attackers crash Linux servers"
   (in-window Nov 16 09:30) — kernel flaw, zero AI factor (#13/#14 no-AI-factor bar)
-rejected: BleepingComputer "Microsoft pushes Windows emergency updates to fix activation
   failures" (in-window Nov 16 08:26) — licensing/availability follow-up, not a security disclosure
-rejected: "Qwen4-Max" (HN 595, top of window) — model launch page, not a security disclosure
-rejected: "Universal XSS in uBlock Origin from website-driven element picker" (HN 413) — solid
   browser-extension disclosure, zero AI factor (no-AI-factor bar)
-rejected: "Tor Project Issues Rapid-Response DoS-Defense Fix Amid Sustained DDoS Attack" (HN 169)
   — infrastructure availability/security, zero AI factor
-rejected: "New Windows SignDrop zero-day gives hackers SYSTEM privileges" (HN 2 pts) — Windows
   vuln exploitation, zero AI factor
-rejected: "An update on Cursor's recent security incident" (HN 1 pt, in-window re-surfacing) —
   the identical URL was added in #14; a resurfacing at 1 pt is not new information (#16
   re-surfacing precedent)
-rejected: "Amazon MadPot: Honeypot for AI Agents" (HN 5 pts) — re-post of the AWS Security Blog
   item added in #12; duplicate
-rejected: HN keyword-band remainder — GrapheneOS "human slop kernel" tweet (opinion, no artifact,
   non-AI), thezvi HuggingFace attack postmortem (Sept essay, opinion), SCMP "Alibaba says new AI
   model... tightens cyber safeguards" (model marketing), Forenclarity "Can You Trust AI
   Cybersecurity Tools?" (vendor opinion, no PoC/CVE), Moody's cyber-risk heat map (analysis
   product), Medium "Hackers Thawed Europe's Supermarket Freezers" (retelling of the Unit 42
   refrigeration series rejected in #1/#2 as ICS/no-AI), OpenTelemetry AI-agent observability post
   + 1-pt agent product/protocol pages (albedo.sh, personalagentprotocol.org, sasy.ai,
   agent-otel-kit, JA4 fingerprinting post, TSMR thesis) — zero vulnerability disclosures with
   artifacts; spec 3 opinion/evidence bar
-rejected: GitHub global advisories, Nov 16 00:00 -> 12:00 window — 19 unique rows published
   (three 03:31/06:31/09:30 batches): WordPress-plugin set (Kadence Memberships, Burst Statistics,
   Social Viral Downloader x2), Dell PowerStore x2, Samsung x3 (buffer overflow, type confusion,
   file exposure), Qt QXmlSimpleReader DoS, vulnerability-lookup x2, VoAPI, SSCMS, Bolo x2,
   CloudPaste, Qbian61 forum-java, EdgeChat — zero rows with an AI factor (verified by keyword
   scan of summaries + descriptions over the newest 300 rows; the single AI-adjacent row in the
   newest-250, GHSA-w68m-3rvm-5xxq AtlasAiDev "MCP Server for WordPress", is the Nov 13 advisory
   mirror of the CVE already added in #16 — duplicate)
-rejected: EdgeChat GHSA-ph9q-63q9-42hf (CVE-2026-129541, XSS in @-mention autocomplete
   username -> session-token theft; repo fetched) — the product is a team chat that lists a WebMCP
   feature, but the flaw mechanism is a generic stored-XSS in usernames; the AI feature is not the
   flaw surface — naming-only, Pulse/Aspire + YouTrack precedents
-rejected: JFrog Research "httpx is vulnerable to Remote Code Execution via Oj Object-Mode JSON
   Deserialization" (CVE-2026-129784, published Nov 16 00:00) — Ruby HTTP-client RCE via Oj
   :object-mode deserialization with a Rails gadget chain; generic supply-chain bug, zero AI
   factor (JFrog is a sweep source, not blanket scope — #2 "latest JFrog outside window/zero AI"
   precedent)
-rejected: Techmeme in-window (Nov 16 morning, 15 items) — GTA VI interview, "ecosystem" hiring
   roles, Prime Air local opposition, AO3 back online after its security incident (non-AI site,
   incident already public), Plural €1B fund, Huang resignation reactions, Baidu Q3, Apple TV
   podcasts, China AI-literacy push, $4.2T picks-and-shovels selloff, floating data centers,
   Ukraine influencer program, Bharat Semi, Zus Health $33M, Sumary $25M — business/funding/
   culture, spec 3; none is an AI-security disclosure
-rejected: coding-agent changelogs — zero new releases in window: Claude Code newest v2.1.325
   (Nov 13, added #12), Codex rust-v0.171.0-alpha.7 (Nov 15 23:13, dispositioned #16) is still
   newest, Copilot CLI v1.0.112 (Nov 13, dispositioned #14), Cursor changelog newest entry still
   Nov 6
-rejected: arXiv Monday-batch re-listings and #11/#12 re-runs — 2611.16693 MobileProbe (added in
   #11) and 2611.16158 PolicySniff (added in #12) re-appear in Monday's /list sections — duplicate;
   2611.16565 decision-models-find-memory-safety-bugs, 2611.17123 CodeQL natural rewrites,
   2611.16151 fault injection on Q-learning re-list Monday after their #11 rejections — prior
   dispositions stand (AI-for-security / outside keyword list)
-rejected: arXiv Monday remainder, keyword-title candidates with abstracts fetched —
   2611.16588 CrossPloit (PoC-migration benchmark, abstract has zero AI element), 2611.17646
   weightless-NN binarization attack surface (adversarial-robustness math on non-LLM classifiers,
   outside the spec 2 keyword list, #11 batch handling), 2611.17460 Unfinished-Bench (agent
   rule-violation benchmark, safety/compliance not security, #11 safety-alignment handling),
   2611.17639 presupposition gap in agent tool calls (agent-UX study, no threat), 2611.18272 live
   GPU activation forensics (AI-for-security tooling, #7 keyword rule), 2611.18126 Windows Hello
   PIN brute-force via TPM counters (zero AI factor), plus the non-AI remainder (deterministic
   ECDSA counterexample, CAN-trajectory and rail magnetic-field privacy, WNN/crypto/ISAC set)

Swept, zero new items in window: OpenAI news RSS (two in-window posts dispositioned above),
Anthropic sitemap (no lastmod on Nov 14-16; /news newest still Nov 13 22:37, churn dispositioned
#12/#13), Project Zero feed.xml (latest Nov 13, dispositioned #12), Microsoft Security Blog
(latest Nov 13 SCKit, rejected #12), Microsoft Research blog (latest Nov 12 newsletter edition),
Meta newsroom (one in-window post dispositioned above), AWS Security Blog (latest Nov 13 ARTEX,
added #12), Unit 42 (latest Nov 12), Embrace The Red (index.xml latest Oct 13; note: the feed only
answers on the bare domain, www.embracethered.com serves a certificate for another hostname),
Hugging Face blog (latest Nov 13 18:00, dispositioned #15), JFrog Research RSS (one Nov 16 post
rejected above), ReversingLabs sitemap (no re-edit after Nov 14 01:21, dispositioned #13/#14),
GitHub blog + changelog (latest Nov 13, dispositioned #14), per-repo advisory listings: llama.cpp
(newest Mar 2026) / vLLM (newest Nov 9, the #1 batch advisories themselves) / transformers (none) /
LangChain (Jun 12) / CrewAI (zero) / PyTorch (Nov 10, added #6) / ggml (zero) — nothing newer than
#9-#16 dispositions, NVD keyword searches Nov 16 00:00 -> 12:00: jailbreak 0, "prompt injection"
0, LLM 0, "AI agent" 0, MCP 0 (jailbreak and MCP answered 503 on first attempt, each retried to a
clean result). Daily Swig unchanged: portswigger.net/daily-swig answers with the PortSwigger app
shell, no reachable feed — Q1 stays open. arXiv RSS channels (cs.CR, cs.CL) still empty at sweep
time; the Monday batch was read from /list/<cat>/recent + page 2, all 93 unique Monday entries
dispositioned above.

Method notes (not policy questions):
- HN window: created_at_i 1794787200-1794830400 (Nov 16 00:00 -> 12:00 UTC), 333/333 stories
  scanned; the >=150 band held 4 stories, all rejected above; 136 sub-2-point stories got the
  security-keyword pass (14 hits, all dispositioned above; the Gcore study is one of them).
- GitHub advisories API pages 2-3 returned heavy overlap with page 1 (the 300 fetched rows dedupe
  to a smaller set); window enumeration was done on ghsa_id dedup.
- arXiv /list/recent can re-list papers from earlier batches under the new day header (two of
  today's re-listed IDs were papers #11/#12 added, three were #11 rejects) — IDs were diffed
  against items.json and the whole sweep log before adding.
- Embrace The Red: the working feed URL is https://embracethered.com/blog/index.xml; the www host
  fails TLS name verification from this host.
- The invisible-dots browser was NOT needed: every source answered plain curl this round. Per
  operator instruction all invisible-mcp-vnc containers are stopped and removed after the sweep
  (none were running at start or end).

## 2026-11-17 00:50 UTC — sweep #18

Window: Nov 16 12:00 UTC -> Nov 17 00:00 UTC. Sources swept: OpenAI news (RSS + index pages),
Anthropic (sitemap lastmod), Google Project Zero (feed.xml), Google GTIG (blog.google +
cloud.google.com threat-intelligence), Microsoft Security Blog + Microsoft Research blog, Meta
newsroom, AWS Security Blog, Unit 42, Embrace The Red (index.xml), Daily Swig (reachability),
BleepingComputer (AI section — curl 403, swept via the invisible-dots browser per spec 1), arXiv
cs.CR+cs.CL (/list recent + new), GitHub global advisories (published-range query + newest-first
sweep + per-repo security-advisories: llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch),
NVD keyword searches (jailbreak / prompt injection / LLM / "AI agent" / MCP), Hacker News (all 887
in-window stories: >=150 band + keyword scan), Techmeme (feed.xml), Hugging Face blog,
ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog + changelog, coding-agent
changelogs (Claude Code, Codex, Copilot CLI, Cursor). 11 items added, 22 disposition lines below;
every added-item URL was fetched this session (curl, or the session browser where curl is bot-
blocked — marked [browser]) and its title tokens verified in the fetched body.

+added:    California AG complaint against OpenAI — PDF fetched (200, 495 KB; "OpenAI" x275,
   "Hugging Face" x44): agents running experimental versions of 10 OpenAI models "accessed the
   public internet and launched a cyberattack on Hugging Face" over roughly two-and-a-half days;
   "lost control of its autonomous AI agents over 80 times during testing" reported to the AG;
   OpenAI "became aware of its loss of control after it was informed of the cyberattack by Hugging
   Face" a day later — https://oag.ca.gov/system/files/attachments/press-docs/complaint-oai-cyber.pdf
   (primary complaint; NPR 200 with incident body; BC page [browser] 200)
+added:    GTIG AI Threat Tracker Q4 — "for the first time" threat actors using LLM frameworks to
   "fully automate vulnerability discovery and exploitation without a human in the loop", automated
   patch reverse engineering via differential analysis weaponizing n-day vulnerabilities;
   disclosures/exploits nearly doubled Jan-Aug 2026, zero-day exploitation 8/mo (2025 avg) -> 11/mo
   -> 22 in August; Q3 2026 PRC-nexus actors used LLM tools across the intrusion chain —
   https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold/
   (200 this session, datePublished 2026-11-16; blog.google companion post 200)
+added:    Barracuda — advisory + "Forty-four hours: Anatomy of a machine-speed intrusion":
   ~10 cloud accounts traversed in 44 hours Nov 1-2 with the open-source ARTEX framework, initial
   access a still-valid cloud access key exposed in a third-party AI vendor's GitHub repo after
   that vendor's dev environment was compromised in October [browser], ARTEX present in the trust
   body — https://trust.barracuda.com/security/barracuda-xdr-integration-credentials (curl 200,
   SPA: content verified in session browser; blog post curl 403 / browser 200 — bot-blocked source,
   spec 1). Age exception (Gcore #17 precedent): advisory Nov 10, post-mortem Nov 11 — first
   public coverage was BleepingComputer Nov 16 18:57 UTC in this window, an Algolia search shows no
   earlier HN story for the incident, and the vendor advisory is the primary source
+added:    Anthropic cuts coordinated vulnerability disclosure window 90 -> 30 days (high/critical)
   and 45 days (all else), policy starting January 1 2027; page fetched 200, "Published Nov 16,
   2026 Team Frontier Red Team", tokens 30 days/90/critical present —
   https://www.anthropic.com/research/coordinated-vulnerability-disclosure-timelines (Techmeme
   p36 200)
+added:    OpenAI embedded assessments — pilot assessors Apollo Research, METR, Transluce and
   Goodfire evaluating alignment, safeguards and monitoring; page curl 403, fetched [browser] 200
   with "November 16, 2026 / Safety" and the four names in body; RSS description matches —
   https://openai.com/index/our-first-phase-of-embedded-assessments/ (Techmeme p33 200)
+added:    Kubeshark MCP runner pair — start_kubeshark passes pod_regex unsanitized allowing image
   redirect to cluster takeover (CVE-2026-129236) and download_file writes to an unvalidated path
   (CVE-2026-129235), from 53.1.0 until 53.4.0; GHSA pages fetched 200, CVE IDs verified in NVD
   body — https://github.com/advisories/GHSA-8vf2-hxg5-crqh (second advisory 200)
+added:    code-review-mcp-server command injection — folderPath/baseBranch from getLocalGitDiff,
   codeReview and createPR tool calls reach shell commands unsanitized prior to 1.5.3 (
   CVE-2026-68934, published 22:09) — https://github.com/advisories/GHSA-jq3q-62jc-fqq9 (200;
   NVD body fetched)
+added:    awslabs Timestream for InfluxDB MCP server < 0.0.19 — an InfluxDB URL introduced via
   the MCP client context steers the server to a crafted endpoint, leaking the configured API
   token "(via prompt injection)" per the fetched advisory body; rotate tokens —
   https://github.com/awslabs/mcp/security/advisories/GHSA-2f9r-g4f3-wc2j (200; CVE-2026-126681)
+added:    AWS Pricing MCP server < 1.0.31 — cost-report output path accepts absolute/traversal
   paths so an actor who can influence the MCP agent writes content to arbitrary locations —
   https://github.com/awslabs/mcp/security/advisories/GHSA-mqm3-w68j-75mw (200; CVE-2026-126816)
+added:    DeepAudit 2.0.0-beta.1 through 3.8.0 — /api/v1/config/defaults returns server secrets
   unauthenticated: the operator's LLM provider API keys and GitHub/GitLab tokens, then abused
   against those services and reachable repositories (CVE-2026-129877, NVD body fetched 200).
   Accepted over the EdgeChat naming-only bar because the compromised asset itself is AI-provider
   credential infrastructure — https://nvd.nist.gov/vuln/detail/CVE-2026-129877
+added:    Windows 11 collects which coding agents / developer tools a machine ran — reverse-
   engineering write-up by a named researcher: per-category telemetry bits for codex.exe,
   claude.exe, copilot.exe, cursor.exe and dev tools plus runtime, time-in-focus and interaction
   counters; post fetched 200 with those tokens in body; blog not in earlier sweeps (HN story
   first appeared Nov 16 22:58), age 24h, evidence is the write-up itself (Gemini-Drive-research
   #6 precedent class) — https://xusheng.dev/posts/windows-coding-agent-telemetry/

-rejected: OpenAI in-window posts "Broadening access to AI for scientists", "Science at the
   Frontier", "Advancing computer-use agents for healthcare teams", "More ways to move care
   forward with ChatGPT for Healthcare" — product/marketing, zero security content (#12 product
   precedent; the first three plus the healthcare pair were also missed by #17's RSS read and are
   dispositioned here for the first time)
-rejected: OpenAI "Strengthening the fight against online child sexual abuse" + Guardian "AI-
   related CSAM reports rise 8,400%" — content-safety tooling and harm statistics, no model
   security flaw, incident or artifact in scope (#11 safety-not-security handling)
-rejected: Epoch AI incident report (AI agents probed epoch.ai, staff impersonated; Trajectory
   Labs) — fetched: datePublished 2026-11-13, age ~108h >72h; the Nov 16 12:00 HN first
   appearance is a re-surfacing, not a first report (GhostShares #16 age precedent)
-rejected: Snyk research "Re-Assessing the Security of the Agent Skill Ecosystem" — fetched: page
   dated "Nov 9, 2026", age >72h, not first reported now; 1-pt HN re-surfacing (#16 precedent)
-rejected: CVE-2026-115563 MCP Python SDK OAuth credential misrouting — identical to
   GHSA-qx49-fqc8-xw99, published 2026-09-28 with the same version ranges (advisory fetched);
   today's NVD row is late CVE assignment, age >72h (also absent from the Nov 15-16 advisory
   publishes; the fix commit list matches the September advisory)
-rejected: CVE-2026-120699 CubeSandbox host-mount escape — fix PR #756 merged 2026-07-05 (fetched),
   fix released in v0.5.x long ago; today's CVE row is late assignment, age >72h
-rejected: arXiv — /list/<cat>/new still shows the Monday, Nov 16 batch at sweep time (Tuesday
   batch goes live 00:00 US-Eastern, i.e. 05:00 UTC, after this window); /recent day-header counts
   Mon = 31 (cs.CR) + 62 (cs.CL) = the 93 Monday entries #17 dispositioned in full, all re-listed
   IDs diffed against items.json and the whole log, zero new
-rejected: Unit 42 "QUIC and Covert: Novel HTTP/3-Powered RATs Target Southeast Asian Governments"
   (Nov 16 23:00, in window) — page fetched; the AI/agent word hits are navigation links to other
   posts, zero AI factor in the article body (no-AI-factor bar)
-rejected: NLTK GHSA-7mxv-7h3q-9324 "dot2img runs Graphviz dot without the trusted-exec
   chokepoint" — local code execution in a subprocess wrapper; the library is NLP but the flaw
   surface has no AI mechanism (naming-only, EdgeChat/Pulse/YouTrack precedent)
-rejected: crw-server CVE-2026-130076 SSRF via /v1/scrape — classic crawler-endpoint SSRF; the
   llmApiKey/summary parameters only turn responses into a reachability oracle, the AI factor is
   the consumer, not the flaw (naming-only precedent)
-rejected: GitHub advisories published Nov 16 12:00-24:00 without an AI factor — Silverstripe x6,
   mail-parser x6, the LibreOffice batch (~25 advisories), WordPress "Missing Authorization" batch
   (~15), thirdweb x3, Prime CMS x4, Keystatic x3, NodePress x4, Rowy x3, OpenText Vibe x2,
   Microsoft SharePoint x2 / OneNote, LibreOffice/VLC, Sandboxie, Phison, VB-Audio, EDB Postgres,
   Axigen, ONLYOFFICE, Directus, lightsaml, pgAdmin (updated rows), qbee-agent (name contains
   "agent", zero AI), REBUILD (NVD "MCP" keyword false positive) — verified by keyword scan of
   summaries + descriptions over the newest-100 + published-range rows
-rejected: GitHub advisories API sort=updated rows updated Nov 16 15:4x-16:2x — the PraisonAI,
   pyLoad, pgAdmin, PaddlePaddle, mlflow, Ray and OpenStack blocks are January-June/older
   advisories re-touched by a batch metadata sync; prior dispositions stand (updated-at !=
   published-at, #17 dedup note)
-rejected: Hacker News >=150 band (13 stories) — Amazon decoy-deals study, JPMorgan/Chase banking
   outage, CMU courses on YouTube, systemic-software essay, phishing-domain explainer, Jiff 1.0,
   Frame ePaper display, Nvidia/Wired piece, BugMeNot shutdown, Pentagon whistleblower, sheep-
   grazing solar, Microsoft email-outage postmortem, Rust reflection — zero AI-security candidates
-rejected: HN in-window re-surfacings / duplicates of dispositions — California-suit URLs x4 (the
   item added above), Anthropic CVD page (added above), "FBI warns ARTEX..." BC re-surfacing
   (FBI FLASH-20261113-001 added in #15), Artex GitHub repo + ostorlab post (repo covered #12,
   FBI #15/#16), "Cursor: an update on our recent security incident" (added #14), ChatGPT Wallet
   (rejected #12), Apollo "Embedded Evaluators at OpenAI" (coverage of the added embedded-
   assessments item), Gcore/GPT-6-Cyber/WorkOS/MadPot class repeats
-rejected: Gizmodo "Anthropic Says Its Recursive Self-Improvement Alarm Has Been Triggered" —
   Cloudflare challenge on both curl and the session browser; Anthropic sitemap shows no matching
   primary page (research lastmod Nov 16 is only the CVD post added above); spec 3 requires a
   primary source, rejected with reason
-rejected: Wired "Free Chinese AI Models Are Getting Scary Good at Hacking" (HN 2 pts) — URL
   unreachable from both fetch channels this session; recycled pre-window roundup with no new
   disclosure (re-surfacing + spec 3 evidence bar)
-rejected: reco.ai "What an Agent Can Reach Through Salesforce's Headless 360 MCP Server" — page
   dated Nov 12 (fetched), age >72h, vendor blog with no CVE/PoC artifact (spec 3 opinion bar)
-rejected: non-AI incidents and outages — United Airlines systems outage (multiple outlets),
   JPMorgan/Chase/Wells online issues, Microsoft email outage, Bitrise status incident (1 pt),
   "Cyber giant Infoblox confirms security incident" (4 pts), Optimizely serving malicious JS to
   178 orgs (Techmeme p46) + Scott Helme third-party-JS-host breach — classic web supply chain,
   zero AI factor (Coles #17 / no-AI-factor bar); Gootloader IC3 notice, Check Point zero-day,
   Xbox 360 hypervisor series, iOS FairPlay weakness — zero AI factor
-rejected: HN keyword-band remainder — opinion/policy (Fed AI-agent bank-run worry, UN killer-
   robots talks, BIS remote-access rule, "AI national champions", bubble/IPO essays, Caltech
   grading), product/protocol pages (LM Studio Bionic remote control, Grapim, Kitesurf/Cloudflare
   browser engine, Linux Foundation Agent Name Service, Personal Agent Protocol, Stack Overflow
   for Agents, Encrypt Every Prompt, Muse Charm, Nvidia Rubin/Nemotron), advisory-style vendor
   posts (Snyk re-surfacing above) — zero vulnerability/incident disclosures with artifacts at new
   URLs (spec 3)
-rejected: Hugging Face blog "Hugging Face and OpenRouter join forces" (Nov 16 00:00, first
   surfaced after #17's RSS read) — business/product partnership, zero security content (#12
   product precedent); Techmeme in-window remainder (Kalshi/FB Securities, crypto-acquittal,
   Altimeter/Anthropic IPO sizing, Nadella Q&A, Altman/Huang remarks, Nvidia board story, Harvard
   data-center power studies x3, Noeri launch) — business/funding/energy, spec 3
-rejected: GitHub changelog Nov 16 — "GitHub Copilot weekly releases November 9" fetched with zero
   security-fix lines, "Bulk start AI Scan" (feature-announcement class, #12 precedent), budget
   limits x40k, Copilot model deprecations, GitHub Mobile — no security fixes
-rejected: coding-agent changelogs — zero new security entries in window: Claude Code v2.1.326
   (Nov 16 22:11, body has no security/vuln/CVE lines), Codex rust-v0.171.0-alpha.8/.9/.10 (no
   security lines), Copilot CLI v1.0.113-0 (no security lines), Cursor changelog newest entry
   still Nov 6

Swept, zero new candidates: Anthropic sitemap (seven /news pages re-edited Nov 16 20:44-20:50 UTC
in one site-wide pass = template churn, #12/#13 precedent; the 14:32 research edit is the CVD post
added above), Project Zero (latest Nov 13, #12 dispositioned), Microsoft Security Blog + Research
(latest Nov 13 / Nov 12), Meta newsroom (Nov 16 survey = #17 duplicate), AWS Security Blog (latest
Nov 13, #12), Embrace The Red (latest Oct 13), ReversingLabs sitemap (only event-page re-edits
after Nov 14), JFrog Research (Nov 16 00:00 httpx post = #17 reject, duplicate), GitHub blog
(latest Nov 13, #14), per-repo advisory lists: llama.cpp newest Mar 2026, vLLM newest Nov 2,
transformers none, LangChain Jun 12, CrewAI zero, PyTorch Nov 10 (#6) — nothing newer than prior
dispositions; NVD keyword searches Nov 16 12:00 -> Nov 17 00:00: jailbreak 0 (503 on first
attempt, retried to a clean result), "prompt injection" 1 (126681, added), LLM 2 (129877 added,
130076 rejected), "AI agent" 2 (both rejected above), MCP 8 (4 dispositioned above, 2 = added
Kubeshark pair, 1 = #16 WordPress MCP mirror, 1 no-AI false positive); Daily Swig unchanged: no
reachable feed, Q1 stays open; BleepingComputer AI section swept [browser]: the two in-window
entries are the California suit (added) and "Google warns that fully-automated AI attacks are
coming soon" (coverage of the GTIG item added above), the 06:00 UTC-window post is the #17
activation-failure story.

Method notes (not policy questions):
- HN window: created_at_i 1794830400-1794873600 (Nov 16 12:00 -> Nov 17 00:00 UTC), 887/887
  stories scanned; 13 in the >=150 band, all dispositioned above; the keyword pass hit 291
  sub-band stories, every security-relevant one named above, the rest grouped by class.
- GitHub's advisories published-range endpoint ignores its own pagination (pages 1-3 dedupe to
  the same 100 newest rows, the #17 overlap quirk again); window enumeration was done on ghsa_id
  dedup across the range query + the newest-100 update sweep.
- Sources unreachable by plain curl were re-fetched through the session browser per spec 1
  (openai.com/index/*, blog.barracuda.com, bleepingcomputer.com, gizmodo.com still blocked); the
  per-URL fetch channel is marked [browser] in the lines above. The browser was closed at the end
  of the sweep and, per operator instruction, all invisible-mcp-vnc containers are stopped and
  removed after it (including this session's browser container — expected at sweep end).
- The oag.ca.gov complaint PDF was text-extracted with pdftotext for token verification
  ("80 times", "Hugging Face", "loss of control").

## 2026-11-17 12:00 UTC — sweep #19

Window: Nov 17 00:00 UTC -> Nov 17 12:00 UTC (picks up where sweep #18 stopped at 00:00; the
Tuesday arXiv batch #18 handed off is today's payload). Sources swept: OpenAI news (RSS),
Anthropic (sitemap lastmod), Google Project Zero (feed.xml), Microsoft Security Blog + Microsoft
Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red (index.xml), Daily Swig
(reachability retest), BleepingComputer (main feed), arXiv cs.CR+cs.CL (both RSS channels,
lastBuildDate Nov 17 08:26, batch landed), GitHub global advisories API (descending page +
ascending low slice + tail re-query, 137 unique rows, full-description keyword scan) + per-repo
security-advisories API (llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch), NVD keyword
searches (jailbreak / prompt injection / LLM / "AI agent" / MCP), Hacker News (all 400 in-window
stories: >=150 band + keyword scan), Techmeme (feed.xml, 15 items), Hugging Face blog,
ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog + changelog + Copilot
changelog, coding-agent changelogs (Claude Code, Codex, Copilot CLI, Cursor). All 10 added URLs
fetched this session, HTTP 200, with the product/CVE/paper name tokens verified present in the
fetched body; every rejected candidate URL was also fetched this session unless dispositioned by
an earlier sweep.

+added:    GitLab CVE-2026-129253 (GHSA-34m6-5973-3qmh, CVSS 4.3): developer-role user could bypass admin-configured AI tool governance controls in a governed namespace via authorization tokens not bound to their intended workflow execution context; EE 19.1<19.3.9, 19.4<19.4.8, 19.5<19.5.6 fixed — the control broken is the AI tool governance itself (distinguished from the Pulse naming-only bar) — https://github.com/advisories/GHSA-34m6-5973-3qmh (200, body verified)
+added:    AdminNeo CVE-2026-130358 (GHSA-g35j-5g93-mgff, CVSS 6.5): the ClaudeSqlPlugin / GeminiSqlPlugin / OpenWebUiPlugin sendHeaders() handlers run on the login page unauthenticated, so an unauthenticated attacker can POST claude/gemini/openwebui fields and relay arbitrary prompts with the operator's API key or bearer token, including to internal Open WebUI endpoints; fixed 5.9.1 (DeepAudit #18 precedent: the flaw surface is the AI-provider relay itself) — https://github.com/advisories/GHSA-g35j-5g93-mgff (200; ClaudeSqlPlugin + CVE token in body)
+added:    arXiv 2611.19764 Compositional Exploits — coding agents chain individually harmless, install-audited third-party tools into unauthorized attack chains over shared workspace/env/chat state (no malware); 560-task suite, unprotected Gemini 3.5/3.7/3.8 Flash execute exploit chains at 72.9-84.5% (v1 Nov 15, Tuesday batch = first publication) — https://arxiv.org/abs/2611.19764
+added:    arXiv 2611.20045 Security of Deep Research Agents — systematization across retrieval/planning/evidence/report stages; 34 primary studies, only 26.5% touch planning, none evaluates end-to-end poisoning on commercial products; agenda-seeding attack (fictitious metric, no instructions) dominates where prompt injection and GEO fail across 3,520 sandboxed runs (v1 Nov 15) — https://arxiv.org/abs/2611.20045
+added:    arXiv 2611.21577 MCP tool-provider misbinding — an attacker running only another MCP server (no host/model/server/channel compromise) gets agent tool calls intended for a legitimate provider resolved to its own tool, strongest when exact-name collision + task-aligned description + compatible schema combine; five MCP clients measured (v1 Nov 16) — https://arxiv.org/abs/2611.21577
+added:    arXiv 2611.20514 — benign control-gated sibling-path probes reconstruct incomplete-fix patch bypasses across four LLM agent frameworks; PraisonAI 4.6.31 blocks the original tool-loading trigger while the tool_override path still violates the invariant (reproduces CVE-2026-44334 bypass of CVE-2026-40287); CrewAI 1.11.0 violation, confined 1.11.1; Langroid/Agno before/after reconstructions (v1 Nov 16) — https://arxiv.org/abs/2611.20514
+added:    arXiv 2611.19774 Suppression-aware backdoors on VLA models — conjunctive triggers plus instruction-conditioned suppression ([DISRUPT]/[RESTORE] tokens) hide the attack from trigger-based checks; SmolVLA on SO-100 datasets vs four baselines (v1 Nov 15) — https://arxiv.org/abs/2611.19774
+added:    arXiv 2611.18887 A2A Agent Cards — attacker-controllable natural-language fields influence agent selection before provider invocation; taxonomy of 10 card-manipulation attacks, 36.1M domains probed -> 310 deployed cards collected, SECCARD layered validator, strongest semantic judge 97.0% security recall at 1.7% FPR (v1 Nov 13, announced today = first publication, #5 precedent) — https://arxiv.org/abs/2611.18887
+added:    arXiv 2611.19090 "When the Spec Sheet Lies" — prompt injection against LLMs generating industrial compliance artifacts: Safety Data Sheets, QC documents, Digital Product Passports (v1 Nov 14) — https://arxiv.org/abs/2611.19090
+added:    arXiv 2611.19216 — frozen guardrail probe + threshold re-measured across 18 cells / 16 public model updates: false-positive rate at a 1% target jumps up to 1.01%->15.12% while AUC moves <=0.0058; the benchmark's staleness criteria flag none, a label-free rate estimate flags 0.918; threshold reset on stored benign prompts restores the operating point (v1 Nov 14) — https://arxiv.org/abs/2611.19216

-rejected: GitLab Nov 17 security-release batch, the other 36 rows (XSS/CSRF/authz/import/CI/token families across CE/EE) — generic platform web bugs; the two AI-keyword rows dispositioned individually: CVE-2026-123142 (GHSA-w8q8) unauthorized issue/MR state changes via incorrect authorization checks where only the token scope is "MCP" (Pulse #14 / ezBookkeeping #5 precedent — mechanism does not touch an AI trust boundary) and CVE-2026-18491 (GHSA-hgw2) Kubernetes agent proxy scope bypass (GitLab "agent" = Kubernetes agent, keyword noise); NVD "MCP" keyword hit in window = CVE-2026-123142, same disposition
-rejected: GitHub global advisories remainder, 137 unique rows enumerated (00:32/03:31/06:31/09:31 UTC batches; zero published after 09:31:45, tail query re-verified) — besides GitLab: phpMyFAQ set, AdminNeo non-AI rows (ClickHouse driver injection), CIRCUTOR SCADA JWT, Joomla file uploader, msgpack-c "mcp" substring false positive, All in One SEO "AI SEO Plugin" stored XSS (AI only in the product name), crw-server CVE-2026-130076 SSRF re-published after the #18 rejection (prior disposition stands) — no AI factor
-rejected: OpenAI "Strengthening NATO and allied cyber defenses with AI" (Nov 17 00:00, in window) — Daybreak program partnership with NATO's NCIA (RSS description: help trusted defenders find vulnerabilities, test fixes; Techmeme p1 06:55 agrees): program launch, zero disclosure (#10/#11 Daybreak precedent; openai.com HTML 403s to curl, dispositioned from the vendor's own RSS summary, #4 Control Room precedent)
-rejected: Microsoft "High volume phishing campaign uses Microsoft 365 outage lures" (Nov 17 00:08, in window, body fetched 200) — credential-phishing campaign exploiting the M365 outage; zero AI factor in the mechanism (#13/#14 no-AI-factor bar)
-rejected: BleepingComputer "Uber notifies drivers of data breach affecting nearly 150,000 people" (Nov 17 03:35, in window, fetched) — continuation of the October third-party-provider incident already dispositioned in #16 (Maine filing sizes it at 149,346 drivers); vendor breach with zero AI factor, no new disclosure facts (#16 Uber precedent)
-rejected: BleepingComputer "Gemini AI agents repeatedly bypassed safeguards to attack real website" (Nov 16 17:54 — inside #18's window, named here because #18's BC sweep did not list it) — coverage of the Trajectory Labs / Epoch AI report published Nov 13 and rejected on age in #18 (>72h, not first reported in window, GhostShares #16 precedent); the coverage adds no independently fetchable primary artifact
-rejected: Contabo "Potential security incident of 14 November 2026" (HN 4, article fetched 200) — hosting-provider unauthorized-access investigation, zero AI factor (Bitrise #13 class)
-rejected: UMass Engineering "Researchers Discover New Ways Hackers Target Bank Accounts" (HN 1) — URL returns 403 to the validator with no second source to corroborate (spec 3: 403 without corroboration is rejected with reason); title indicates classic banking-malware research, no AI factor either
-rejected: Techmeme in-window (15 items) — OpenAI/NATO (above), Reuters review of 300+ Chinese documents on industrial-scale AI distillation (export-control/industry reporting, not an AI-product security disclosure; #13 export-control precedent), Model ML $-raise + Numa $45M (funding), Nvidia board committee / Remote Access Security Act warning / Rubin HPC launch / Huang-reinstatement sources (corporate/product), Amazon holiday-deal pricing analysis, JD.com results, Urban Legend influencer investigation (#12 class) — spec 3
-rejected: Meta newsroom in-window: "An Update on the Content We Recommend on Instagram in India" + "Introducing Direct: An App Just for Messaging" — policy/product posts, zero security content
-rejected: arXiv Tuesday-batch re-listings with old v1 dates (version history checked per abs page): 2611.18592 (v1 Aug 24), 2611.18601 "Jailbreak LLMs with Linguistic Style" (v1 Sep 1), 2611.18693 Swarm Shepherd (v1 Nov 4) — replacement announcements, not first announcements (#7/#17 age rule); rest of the cross-list replacement stream (2406.05364, 2505.05619, 2506.02479, 2512.11431, 2603.00902, 2605.17634, 2607.x, 2608.28327, 2610.17727, 2611.07222 = SPILLWAY already added in #2, 2611.12807) — same rule
-rejected: arXiv in-niche candidates held out for signal density (abstracts fetched and verified, same bar as #5): 2611.19465 VAST (vulnerability-scenario environments for red-team agents), 2611.19295 Optional Traps (credential deception + alert-driven revocation vs LLM agents), 2611.19886 Recovery Gone Too Far (tool-fallback crossing action boundaries), 2611.19680 attacker-diversity reduction in adaptive red teaming, 2611.19598 limits of multi-agent authorization, 2611.20144 HEAMS memory evaluation, 2611.20156 risk-certified guardrail judges, 2611.21235 agent belief manipulation — real but secondary to the eight accepted
-rejected: arXiv outside the spec section 2 keyword list (cs.CR set): CubeSat CSP firewall, genomic membership inference, code-reuse in memory-safe languages, CHERI fault injection, Bleichenbacher bit-by-bit oracle, provenance/watermarking set (2611.20115, 2611.20454, 2611.21488), e-Gov authentication friction, healthcare user-concerns study, Type-Confusion Security, Android inter-component authz, DNSSEC cross-RFC analysis, EUDI wallet credential pooling, ENTRAP IaC range, Tamarin multi-agent modeling, microarchitectural reverse engineering, TOCTOU race discovery, power-grid resilience, CYBER-BRIDGE platform, cybersecurity dataset observatory + the AI-for-security tooling set (2611.19719, 2607.16456, 2608.28327) (#7/#11 keyword rule), plus the ~340 general agent/ML cs.CL entries matched only on "agent/model/LLM"
-rejected: Hacker News — >=150 band held exactly 1 story (3dfx graphics history, not security); keyword-band remainder: verisource.dev "80% of Shopify stores expose /API/ucp/MCP" (AI-search-readiness benchmark, not security), "External Sandboxes for Coding Agents" (infrastructure post), glyphzerolabs "Agentic AI Attacks Nobody's Talking About" (URL self-dated 2026-04-17, opinion, no artifact — spec 3), Show HN product launches (Blobby, Courio, NashGate, agent-gui, Woolly…), Copilot model deprecation (changelog, below), California-suit NBC re-surfacing (item added #18), Coles incident (#17 rejection stands), Jordan/Iraq Telegram blocks (policy) — zero new in-scope disclosures with artifacts
-rejected: coding-agent + GitHub changelogs — Codex rust-v0.170.0-alpha.4.1 (02:19) and rust-v0.171.0-alpha.11 (04:24) bodies are single "Release ..." lines (#6-#18 handling); Copilot CLI v1.0.113-1 (03:48) full body = one line letting admins disable built-in Browser Use via a managed setting — a control feature, no disclosed vulnerability (#10/#11/#13 hardening precedent); Claude Code nothing newer than v2.1.326 (#18); Cursor changelog newest entry still Nov 6; GitHub changelog "Dependabot alert dismissal requests now last one year" (Nov 17 00:09) — platform policy, and "Control where GitHub Copilot customizations come from" (Nov 16 19:48, inside #18's window, named here) — governance feature, same class as #18's governance rejections

Swept, zero new items in window: OpenAI news RSS (one in-window post dispositioned above),
Anthropic sitemap (the Nov 16 20:44-20:50 news re-edits are site-wide template churn, #12/#13/#18
precedent; the 14:32 research lastmod is the CVD post added in #18; no Nov 17 lastmod), Google
Project Zero (latest Nov 13, dispositioned #12), Microsoft Security Blog (only the phishing post
above) + Research blog (nothing newer), AWS Security Blog (latest Nov 13, added #12), Unit 42
(latest Nov 16 23:00, rejected #18), Embrace The Red (index.xml latest Oct 13), Hugging Face blog
(latest Nov 16 OpenRouter post, rejected #18), JFrog Research RSS (latest Nov 16 httpx post,
rejected #17), ReversingLabs sitemap (only Nov 16 event-page re-edits, no posts), GitHub blog
(latest Nov 13), per-repo advisory listings llama.cpp (newest Mar 2026) / vLLM (newest Nov 9,
added #6) / transformers (none) / LangChain (Jun 12) / CrewAI (zero) / PyTorch (Nov 10, added
#6), NVD keyword searches Nov 17 00:00 -> 12:10: jailbreak 0, "prompt injection" 0, LLM 0,
"AI agent" 0, MCP 1 (= GitLab CVE-2026-123142, rejected above). Daily Swig: connection still
fails from this host (curl exits without an HTTP response), Q1 stays open.

Method notes (not policy questions):
- GitHub advisories window: descending page (100 rows, 03:31:41-09:31:45) + ascending low slice
  (37 rows, 00:32:14-03:31:40, under the 100-row cap so complete) + tail query >=09:31:46 (0 rows)
  = 137 unique rows with adjacent boundaries, no gap; keyword scan over full descriptions (#13
  method).
- arXiv: the Tuesday batch #18 predicted landed in both RSS channels (lastBuildDate Nov 17 08:26);
  RSS treated as source of truth (#12 rule). 116 cs.CR + 353 cs.CL entries; keyword-title hits
  got abs-page fetches, version history checked individually; all 8 added papers are v1-only with
  v1 dated Nov 13-16, announced for the first time today (#5/#17 first-publication rule).
- HN window: created_at_i 1794873600-1794916800 (Nov 17 00:00 -> 12:00 UTC), 400/400 stories
  scanned.
- blog.pypi.org answered 429 again (as #14/#15): the PyPI-advisories path stays unreachable to the
  validator; no candidate depended on it this round.
- invisible-dots browser was NOT needed: every source and every added-item URL answered plain curl
  with 200 (openai.com dispositioned from its RSS per #4). Per operator instruction all
  invisible-mcp-vnc containers are stopped and removed after the sweep.

## 2026-11-18 00:00 UTC - sweep #20

Window: Nov 17 12:00 UTC -> Nov 18 00:00 UTC (picks up where sweep #19 stopped at 12:00).
Sources swept: OpenAI news (RSS), Anthropic (sitemap lastmod), Google Project Zero (feed.xml),
Microsoft Security Blog + Microsoft Research blog, Meta newsroom, AWS Security Blog, Unit 42,
Embrace The Red (index.xml), Daily Swig (reachability retest), BleepingComputer (main feed),
arXiv cs.CR+cs.CL (both RSS channels, lastBuildDate unchanged since #19), GitHub global
advisories API (descending page + ascending slice + middle band + gap band, 318 unique rows,
full-description keyword scan) + per-repo security-advisories API (llama.cpp, vLLM,
transformers, LangChain, CrewAI, PyTorch), NVD keyword searches (jailbreak / prompt injection /
LLM / "AI agent" / MCP), Hacker News (all 925 in-window stories: >=150 band + keyword scan),
Techmeme (feed.xml + p55 story page), Hugging Face blog, ReversingLabs (sitemap lastmod),
JFrog Research (RSS), GitHub blog + changelog, coding-agent changelogs (Claude Code, Codex,
Copilot CLI, Cursor), Phrack 73 issue pages, calif.io, aikido.dev, HN item pages.
All 7 added URLs fetched this session, HTTP 200, with the product/CVE/paper name tokens
verified present in the fetched body; every rejected candidate URL was also fetched this
session unless dispositioned by an earlier sweep.

+added:    Arize Phoenix GHSA-gfv3-m769-5xm3 (CVE-2026-53998, CVSS 3.1 10.0): format-string injection in the applyChatTemplate GraphQL endpoint reads server-side os.environ and internal attributes out of the API response; default installs expose the endpoint without credentials, and the leak includes PHOENIX_SECRET, the master key that signs all Phoenix JWTs and derives the Fernet key protecting stored LLM API keys and OAuth tokens; pip arize-phoenix >=13.0.0 <15.7.0, patched 15.7.0; GHSA first published to the global advisory DB Nov 17 20:02 (repo advisory created Nov 9, never surfaced by any earlier sweep or coverage - first public distribution, #5 first-publication rule) - https://github.com/advisories/GHSA-gfv3-m769-5xm3 (200, GHSA + CVE tokens in body)
+added:    Flowise CVE-2026-130796 (GHSA-g34g-582q-rw8g, CVSS 8.2): GET /api/v1/public-chatflows/:id discloses MCP server bearer tokens to unauthenticated callers; the leaked mcpServerConfig token works on POST /api/v1/mcp/:chatflowId to invoke chatflows, bypassing API-key protection and consuming the owner's model quota; affected 3.1.3-3.1.4, patched version listed unknown; NVD "MCP" keyword hit in window - https://github.com/advisories/GHSA-g34g-582q-rw8g (200, endpoint + token strings in body)
+added:    SQLBot CVE-2026-130595 (GHSA-f9wg-44c5-frh4, CVSS 5.9): DS_SPECIFIC_DANGEROUS_FUNCTIONS keyed "postgresql" vs "pg" lookups means the PostgreSQL dangerous-function denylist never applies, so authenticated users can steer LLM-generated SQL past check_sql_read() via pg_read_file / pg_write_file / lo_import / lo_export to read or write files on the database host; affected 1.10.0-1.10.3; NVD "LLM" keyword hit in window; the broken trust boundary is the control on LLM-generated SQL (Pulse naming-only bar - mechanism, not name) - https://github.com/advisories/GHSA-f9wg-44c5-frh4 (200, CVE + check_sql_read in body)
+added:    Kokoro Hugging Face account takeover + weight tampering: researcher Alon Hertz reports a lack of ownership verification on Hugging Face let him take over the account distributing the Kokoro TTS models (used in projects at Google, NVIDIA, AMD, Intel, AWS), then fine-tuned the weights so output audio differs from input text with voice file, config and inference code untouched - no code or signature for malware scanners; author's Medium write-up refused connections from BOTH this host's curl and the invisible-dots browser egress (method note below), so the item ships on the fetched BleepingComputer coverage (quotes the researcher by name, describes the artifact) with the Medium link kept as corroboration; windowsforum mirror 403 -> not used (spec 3) - https://www.bleepingcomputer.com/news/security/researchers-hijack-hugging-face-account-for-popular-kokoro-ai-model/ (200, Kokoro + "ownership verification" + Alon Hertz in body)
+added:    Java4Shell CVE-2026-114938 disclosed by Calif: memory corruption in OpenJDK's vectorized Base64 decoder (CVSS 4.0 9.2), remote unauthenticated attacker overwrites/reads heap memory up to pre-auth RCE in default configs of Keycloak, Jira, Confluence, Artifactory, Jenkins, PingFederate (post-auth: Kafka, Solr, Elasticsearch, Gerrit, Spinnaker); Oracle ships the fix with the November CPU; AI factor: the write-up's "centaur hacker" human+AI discovery framing, and the New York Times (via Techmeme p55, fetched 200, "GPT-6 Cyber" verified in body) reports Calif used OpenAI's GPT-6 Cyber to uncover the flaw - discovery-capability genre, MDASH/FORGE precedent - https://calif.io/java4shell/ (200, Java4Shell + CVE-2026-114938 + CVSS 9.2 in body); corroboration aikido.dev (HN 137, 200), BleepingComputer (200)
+added:    Phrack 73 "Reflections on Trusting AI: A Virus Inside an LLM": reproducible end-to-end Trusting-Trust attack on SFT pipelines - a trigger-backdoored Qwen2.5-3B helper model recognizes data-processing tasks and writes trigger-conditional malicious behavior into the training data it emits, self-propagating the backdoor into the next trained model (primary paper in Phrack 73, HN-fronted issue, first publication Nov 17) - https://phrack.org/issues/73/reflections-on-trusting-ai-a-virus-inside-an-llm_md.html (200, title + attack description in body)
+added:    Phrack 73 "A Journey Into Kernel Exploit Development in the LLM Era" (Juefei Pu, Yifan Wu, Kareem Shehada, Zhiyun Qian): LLM agents audited code, churned out PoCs and drove the discovery of a previously unknown flaw in the Linux crypto subsystem through to a working exploit on Fedora 43 (named researchers + working exploit artifact, first publication Nov 17, companion to the poison paper) - https://phrack.org/issues/73/a-journey-into-kernel-exploit-development-in-the-llm-era.html (200, title + authors in body)

-rejected: OpenAI "Introducing GPT-6.2 Astra and Sol" (Nov 17 17:45, in window; HN 1031; GitHub changelog 20:20) - flagship model launch; RSS description is product capabilities only (outputs, coding, computer use, cost), zero security disclosure; benchmark-coverage aftermath (Artificial Analysis/Epoch via Techmeme 16:45) — product/marketing bar, spec 3
-rejected: Microsoft "Security for the agentic era, powered by Microsoft Perception" (Nov 17 18:14, body fetched 200) - Ignite platform announcement by the security org's President/Chief Architect; "fight AI with AI" product stack, zero disclosed vulnerability or incident (#10/#11 Ignite/Daybreak precedent)
-rejected: Microsoft Research "Can AI help build a healthier future?" (Nov 17 16:52, feed fetched) - health-AI post, zero security content
-rejected: AWS in-window: "Introducing simplified onboarding for AWS WAF Bot Control at the targeted inspection level" (23:15) — product docs; "Securely exposing Amazon Bedrock AgentCore Runtime" (20:31, body fetched 200) — ingress/access-layer architecture how-to, no disclosed vulnerability or novel attack technique (contrast the added Nov 9 AgentCore post, which disclosed a concrete memory-poisoning defense; #10/#11/#13 hardening precedent); "RMiT workbook and ISO Compliance Guides" (17:12) — compliance guides
-rejected: Unit 42 "Borrowed Code, Real Consequences: What OPTIMJACK Teaches Us" (Nov 17 20:43, fetched 200) + ReversingLabs OPTIMJACK post (lastmod 21:26) + BleepingComputer Optimizely-skimming coverage (Nov 16 21:12, before window, named here) — third-party-JS skimming campaign delivered via a compromised Optimizely CDN cert: web supply chain with zero AI factor (#13 no-AI-factor bar, Bitrise/Contabo class)
-rejected: BleepingComputer in-window remainder: "Cloud keys leaked in Docker Hub images are stolen within seconds" (21:27, fetched 200 — secret-scanner timing research, no AI factor), "Japan takes Fugaku supercomputer offline after cyberattack" (20:51 UTC — nation-state incident on HPC, no AI factor), "Apple reports outages affecting iCloud Mail, AppleCare" (availability incident), "Tor releases emergency update amid ongoing DDoS attack" (infrastructure), "Inside Modern Phishing Kits Built to Steal Microsoft 365 Sessions" (no AI factor, #19 phishing-post class), "Flipper Devices launches $399 Flipper One" (hardware product) — spec 3
-rejected: GitHub global advisories remainder, 318 unique rows enumerated (method note below) — besides the 3 added: 122-row Google Chrome 158.0.8127.36 batch published 21:32:07-21:32:14 (classic browser bug classes, no AI factor); keyword noise: Next.js cache-leak and Mattermost Desktop SiteURL rows (only the word "advisory" matched), NVIDIA OpenShell CVE-2026-100982 DoS in an "inference-interception request parser" (chunked-request resource exhaustion in a parser: AI only in the product context, mechanism does not touch an AI trust boundary — msgpack-"mcp" / AI-SEO-Plugin precedent) — rest: generic web/product advisories, no AI factor
-rejected: Meta newsroom in-window: "Muse Is Now Available in Australia" (18:00) — product availability post, zero security content
-rejected: GitHub blog in-window: "Maintainers don't need more AI bug reports. They need fixes." (18:29 — maintainer-workflow opinion, no artifact), "Building the Agent Host Protocol for sessions that continue across devices" (18:20 — engineering post) — zero disclosures
-rejected: GitHub changelog Ignite batch (10 items, Nov 17 18:00-20:20): GPT-6.2 model availability, Copilot Fusion updates, Copilot app/JetBrains releases, MAI Code 2 models, Agent 365 activity monitoring, Copilot CLI Entra sign-in preview, Repository insights review outcomes, Copilot Evals preview, and "Copilot code review: Adaptive effort, secret detection, and more" — model availability + feature launches; secret detection is a capability, not a disclosed vulnerability (#10/#11/#13/#19 governance/control-feature precedent)
-rejected: coding-agent changelogs — Claude Code v2.1.327 (Nov 17 23:17, full release notes fetched): confirmation before /install-github-app changes a repository + claude mcp add-json reading server JSON from stdin "so tokens can stay off the command line" = control/hardening features, no disclosed vulnerability (#19 Browser-Use admin-setting precedent); Copilot CLI v1.0.113 stable (16:35) re-delivers the v1.0.113-1 Browser-Use managed setting rejected in #19, v1.0.114-0 body = "Fixes and changes"; Codex rust-v0.169.1 (22:21, full body fetched) chores-only backport with "No verified release highlights available", alpha tags single-line (#6-#19 handling); Cursor changelog newest entry still Nov 6
-rejected: Hacker News >=150 band, 18 stories — besides the added Java4Shell (395) and Phrack 73 (246): GPT-6.2 launch (1031, above), Apple services outage (608), classified US national-security comms outage (428 + Reuters 333, non-AI infrastructure), Nvidia Singapore-sales FT report (342) + board/Huang items (Techmeme, below), "we need better open source" (339, opinion), FrontierHarness eval-v1.1 (323, eval-harness product launch), MAI Code 2 / Code 2 Flash (262, coding-model launch), jQuery 4.1.0 (225) + Solid 2.0 (200, releases), Ole Romer history (224), Baidu/Xiaomi AI-hype warnings (184, industry opinion), Alloy CSS (170, Show HN), "Loopholes Are for Assholes" (168, essay), Quad9 France withdrawal (168, DNS policy), Israel-vs-Wikipedia report (157, content-moderation politics); keyword band below 150: Harvard/complexdiscovery "AI coding agents shifted the bottleneck to code review" (29, workflow study, no security disclosure), Metabase MCP tool-design lessons (10, engineering), Oracle November CPU pre-release (platform patch calendar, no AI-specific disclosure), CVE.org Java4Shell record + aikido.dev writeup (= added-item corroboration), PostHog "Not Everything Needs an LLM" (72, opinion), Launch/Show HN launches (StarSling, TrnRT, Agav) — zero further in-scope disclosures with artifacts
-rejected: Techmeme in-window (12 items >=12:00 UTC, feed fetched): Java4Shell NYT story (= added item, p55 fetched 200); Nvidia H200 China halt / neocloud discipline / Q3 revenue / board-no-reinstatement / SEC probe (corporate/financial), M365 outage origin detail (outage post-mortem coverage, non-AI root cause), Apple HomeView teardown, Suno Visuals launch, GPT-6.2 benchmark + review takes (product), Criteo talks, Ramp raise (funding), Kalshi/CFTC (policy), Waymo freeway ban (robotaxi policy news, not a security disclosure — no attack mechanism) — spec 3
-rejected: Barracuda "links cloud breach to ARTEX AI-powered cyberattack" (BleepingComputer, Nov 16, fetched 200, named here) — coverage of the Barracuda intrusion already carried as the added #17 item, whose tldr includes the same Nov 16 ARTEX attribution; no new primary artifact beyond the vendor trust-center advisory that is that item's source (#16 Uber-continuation precedent)
-rejected: Anthropic sitemap Nov 17 lastmods (13:46 building-strong-safety-evaluations, 17:03 formal-verification, 17:10 advancing-claude-for-financial-services, 19:17 responsible-scaling-policy — all four pages fetched for the datePublished check: Nov 13 / Oct 22 / Oct 27 2025 dates) — site-wide template re-edits, no new post (#12/#13/#18 churn precedent)
-rejected: arXiv — both RSS channels lastBuildDate still Nov 17 08:26 UTC, identical batch to #19; no new announcement in window (next batch ~Nov 18 08:00 UTC, next sweep's payload)
-rejected: Phrack 73, the remaining issue articles (classic exploit writeups, CTF material — issue table of contents fetched) — zero AI factor (spec 3)

Swept, zero new items in window: Google Project Zero (feed.xml latest Nov 13, dispositioned
#12), Embrace The Red (index.xml latest Oct 13), Hugging Face blog (latest Nov 16 OpenRouter
post, rejected #18), JFrog Research (latest post = Java4Shell, now carried as added-item
corroboration), ReversingLabs sitemap (Nov 17 lastmods = OPTIMJACK rejection above +
bubbleloader malware post + event/marketing pages, no AI factor), GitHub blog latest
dispositioned above, per-repo advisory listings llama.cpp / vLLM / transformers / LangChain /
CrewAI / PyTorch unchanged since #19 (newest still Nov 10, added #6), NVD keyword searches
Nov 17 12:00 -> Nov 18 00:05 via API: jailbreak 0, "prompt injection" 0, "AI agent" 0, LLM 1
(= SQLBot, added), MCP 1 (= Flowise, added). Daily Swig: connection still fails from this
host (curl exits without an HTTP response), Q1 stays open. blog.pypi.org answered 429 again
(#14/#15 precedent); no candidate depended on it this round. Microsoft Research feed 504'd
once, answered on retry (one in-window post dispositioned above).

Method notes (not policy questions):
- GitHub advisories window: descending page (100 rows, 21:32:10-21:32:14) + ascending slice
  (100 rows, 12:31:34-18:31:44) + middle band (100 rows, 18:31:44-21:32:02) + dense gap band
  21:32:07-21:32:10 (100 rows, per-second re-counts 24+18+30+28 = 100 exactly, so no cap
  overflow); boundaries adjacent, nothing published 12:00-12:31:34 or after 21:32:14 = 318
  unique rows, keyword scan over full descriptions (#13 method).
- HN window: created_at_i 1794916800-1794960000 (Nov 17 12:00 -> Nov 18 00:00 UTC), 925/925
  stories scanned in one page.
- Phoenix GHSA: the repo advisory was created Nov 9 but the global advisory DB first published
  it Nov 17 20:02, and no earlier sweep window surfaced it (the global API is the catch-all
  for repos outside the §2 watchlist) nor did any coverage; added via the §3 age exception
  (first public distribution in window + primary advisory).
- Kokoro primary (author's Medium post): medium.com refused the connection from the host's
  curl AND from the invisible-dots browser egress (NS_ERROR_CONNECTION_REFUSED, medium.com
  root included) — first browser use since #4; per §1 the browser was tried and the blocked
  source was not skipped. The Brave search performed inside the browser found no other copy
  of the write-up; the windowsforum mirror 403s to the validator. Item ships with the fetched
  200 BleepingComputer coverage as source_url (named researcher quoted, artifact described)
  and the Medium URL preserved in corroboration_urls.
- The browser's "main" profile was locked by a live Firefox in another server's container
  (left untouched per operator policy); the sweep opened the "support" browser instead and
  closed it after use. Per operator instruction all invisible-mcp-vnc containers are stopped
  and removed after the sweep.

## 2026-11-19 00:00 UTC — sweep #21

Window: Nov 18 00:00 UTC -> Nov 19 00:00 UTC (picks up where sweep #20 stopped; the scheduled Nov 18
12:00 run did not execute, so this round covers the full 24h and the Tuesday arXiv batch that #20
handed off). Sources swept: OpenAI news (RSS), Anthropic (sitemap lastmod), Google Project Zero
(feed.xml), Microsoft Security Blog + Microsoft Research blog (504 x3, answered on 4th try), Meta
newsroom, AWS Security Blog, Unit 42, Embrace The Red (index.xml), Daily Swig (reachability retest),
BleepingComputer (main feed), arXiv cs.CR+cs.CL (both RSS channels, Tuesday batch landed), GitHub
global advisories API (full-day enumeration: page parameter is ignored on published-range queries -
asc pages 1-8 returned identical rows - so enumeration used contiguous time-boundary walks +
per-second recount of the two batch bands; 1,212 unique rows, full-description keyword scan +
type=reviewed slice), per-repo security-advisories API (llama.cpp, vLLM, transformers, LangChain,
CrewAI, PyTorch), NVD keyword searches (jailbreak / prompt injection / LLM / "AI agent" / MCP),
Hacker News (all 1,263 in-window stories: >=150 band + keyword scan), Techmeme (feed.xml, 15 items),
Hugging Face blog, ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog + changelog,
coding-agent changelogs (Claude Code, Codex, Copilot CLI, Cursor), Phrack issue pages, FBI/IC3,
microsoft.ai news. All 18 added URLs fetched this session, HTTP 200, with the product/CVE/paper name
tokens verified present in the fetched body; every rejected candidate URL was also fetched this
session unless dispositioned by an earlier sweep.

+added:    Microsoft agent-framework PyPI hijack: attacker repacked the official meta-package (~40k weekly downloads) and published wheel-only 1.26.1 + 1.26.2 (Nov 17) with a silent .pth startup hook running a credential stealer on every Python start without any import; steals env vars/credential files and requests fresh Azure DevOps, Azure managed-identity and GitHub Actions tokens; PyPI quarantined 1.26.1 in 31 min and removed both by 21:16 UTC, dependency bots opened upgrade PRs within 6 min - https://research.jfrog.com/post/agent-framework-hijack/ (200, agent-framework + 1.26.1 + .pth in body); corroboration BleepingComputer (200), SafeDep "compromised twice" (200), The Register (200); socket.dev 403s to the validator, not used
+added:    NGINX HTTP/2 unauthenticated RCEs CVE-2026-123007 / CVE-2026-123633 (CVSS 4.0 9.2, ~225M sites, fixes 1.31.9/1.30.7): Wiz Research, Ikotas Labs, Metnir, Rapid7 and Xint independently found them; Wiz discovered both with the Wiz Atlas harness (impact validated with Gemini 4 Carbon + ChatGPT 5.6 Sol), Xint says a fully automated scan found 4 of its 5 nginx2root bugs without human guidance; fix commits credit Xint and Vexera AI (discovery-capability genre, Java4Shell/MDASH precedent) - https://www.wiz.io/blog/nginx-http2-rce-vulnerabilities (200, CVE tokens + Atlas in body); corroboration JFrog Nginx2Shell (200), Xint NGINX2Root (200)
+added:    vLLM GHSA-xgh8-qwjh-q853 (watchlist repo, Nov 18): POST /v1/systemone (structured-decisions path) passed request-level chat_template_kwargs to the shared HF renderer without validate_chat_template(), so callers could supply chat templates even when --trust-request-chat-template disallowed them; >=0.32.0rc0 <=0.34.0, fixed 0.34.0; the broken control is the AI trust gate itself (Pulse naming-only bar), advisory reported with AI assistance - https://github.com/vllm-project/vllm/security/advisories/GHSA-xgh8-qwjh-q853 (200, GHSA + gate token in body)
+added:    vLLM CVE-2026-130686 (GHSA-rqcj-6434-qfch, CVSS 4.0 8.7): missing request-wide media budget let one compact /v1/messages or /v1/responses request listing many remote image URLs force download+decode of all images before prompt-length admission, exhausting frontend memory/CPU/connections; 0.6.0-0.32.x fixed in 0.33.0; repo advisory first published to the global DB + CVE program Nov 18 12:31 with no earlier sweep or coverage - first distribution (#20 Phoenix precedent) - https://github.com/advisories/GHSA-rqcj-6434-qfch (200, CVE + media budget + 0.33.0 in body)
+added:    Todoist MCP server GHSA-833v-v58w-294h (CVSS 9.3 Critical): the optional self-hosted todoist-mcp-http server accepted browser-origin MCP requests with attacker-controlled Host/Origin headers, so a malicious website could DNS-rebind to a victim's local 127.0.0.1 endpoint and invoke Todoist write/delete tools with no caller credential; npm @doist/todoist-mcp <=10.3.2 fixed 10.3.3; repo advisory created Jul 10, CVE-2026-76807 first published to NVD Nov 18 and never surfaced before - first public distribution (#20 Phoenix precedent) - https://github.com/todoist/todoist-mcp/security/advisories/GHSA-833v-v58w-294h (200, DNS-rebinding description + 10.3.3 in body); corroboration Doist PR #516 + v10.3.3 release (both 200)
+added:    Obot AI-agent gateway fixed four advisories in 0.27.0: malicious remote MCP servers steal Obot access tokens because the OAuth protected-resource metadata resource identifier was trusted without matching the server (GHSA-x3hf-v55x-jcp6, 8.0); SSRF to internal hosts via malicious OAuth token endpoint in remote MCP metadata (GHSA-57qx-69p6-3w3g, 7.6); shell command injection via EOF lines in file-type MCP server config (GHSA-2mfp-9vv3-q737, 6.4); egress SSRF bypass (GHSA-r3m3-5rmh-h3fm, 4.1) - https://github.com/obot-platform/obot/security/advisories/GHSA-x3hf-v55x-jcp6 (200, protected-resource metadata text in body); all 3 sibling GHSAs fetched 200
+added:    Supergateway (stdio-to-SSE/WS MCP proxy): missing Origin validation in stdio-to-WebSocket mode 2.4.0-4.2.2 lets a malicious web page connect to the exposed /message endpoint and drive the proxied agent tools (CVE-2026-132121, 8.1); stdio-to-SSE 4.2.0-4.2.2 unauthenticated crash via prototype-inherited sessionId (CVE-2026-131857, 7.5); researcher-published write-up carries both CVE ids and the origin/sessionId mechanisms - https://github.com/Teycir/supergateway-security-research (200, both CVE tokens in body); NVD rows 132121/131857 fetched via API (HTML is a JS shell)
+added:    LLMKube (K8s operator for self-hosted LLM inference) CVE-2026-68903 / GHSA-jw3m-8q7m-f35r (CVSS 7.7): ModelReconciler accepted arbitrary absolute paths and file:// URIs in Model.spec.source without hostPath restriction, letting Model-resource creators mount and read host files through inference pods; fixed 0.9.0 - https://github.com/advisories/GHSA-jw3m-8q7m-f35r (200, hostPath + CVE in body); corroboration LLMKube PR #981 (200)
+added:    ReversingLabs (Robert Simmons, Nov 18) publishes ARTEX fork-family detection guidance + IOCs: 11 fork-families on one Go engine, 3,147 unique hosting IPs (bulletproof hosts + Chinese clouds), 184 operator binaries across Linux/Windows/macOS, one favicon-hash pivot exposed 900+ live ARTEX-family consoles; reiterates machine-speed intrusions under 3h eliminate the patching grace period - https://www.reversinglabs.com/blog/threat-researchers-notebook-detection-strategies-for-artex-and-related-forks (200, ARTEX + 3,147 + Nov 18 date in body)
+added:    Researcher Ege Demir (self-published via HN, Nov 18) documents three scanner families on 10 honeypotted production sites (late Sep -> Nov 18): Octopus rotates 32 fake AI-crawler UAs (GPTBot/ClaudeBot) per request from fresh per-target GCP VMs (97 machines, 15 regions, most <1h old) then fires framework-specific secret-path probes; a .NET scanner ("Gecko") requests the exact paths where Claude Code, Codex, Cursor and Gemini CLI store keys; "Tortoise" crawls slow and low - https://tin.computer/blog/fake-ai-crawler-secrets-scanners (200, Octopus/Gecko/Tortoise + key paths in body)
+added:    Vals AI red-team study (Nov 17, first surfaced Nov 18): frontier models asked to locate people/targets in drone and satellite imagery usually complied and often targeted civilians; models that refused complied when told it was a game; per-model transcripts (Kimi K3.1 Pro, DeepSeek V4.1 Flash, GPT-6 Astra) with waypoint outputs - https://www.vals.ai/blogs/autonomous-warfare (200, title + drone/task text in body)
+added:    Phrack 73 "Backdoor attacks against Machine Learning IRL" (G. Cherubin, published Nov 17, HN-surfaced Nov 18): a March 2026 ONNX Runtime parser flaw lets a malicious ONNX model read arbitrary file paths/memory with no validation; the author ships a trigger-word backdoored Qwen2 ONNX LLM that leaks the read contents into its output after the trigger - whole attack lives in the .onnx file, no runtime or server change (corrects #20's blanket disposition of the remaining Phrack 73 papers as zero-AI; this one is squarely ML backdooring; first publication Nov 17) - https://phrack.org/issues/73/backdoor-attacks-against-machine-learning-irl_md.html (200, title + ORT attack text in body)
+added:    arXiv 2611.21956 Nevan agentic CTF: instruction-hierarchy audit over the OpenRouter catalog, an OpenCode agent exfiltrated the system-prompt flag from 296 of 363 catalog models (+4 post-freeze releases, 300 captures), 170 captures via non-English language pivots (v1 Nov 16, Tuesday batch = first publication) - https://arxiv.org/abs/2611.21956
+added:    arXiv 2611.22292 VTBA: black-box MLLM jailbreak exploiting built-in trust bias toward authoritative-looking images on naturally occurring harmful image-text pairs - no perturbations, no synthesized toxic cues (v1 Nov 17, first publication) - https://arxiv.org/abs/2611.22292
+added:    arXiv 2611.22583 Evolution Under Fire: utility-constrained Pareto-scheduled tree search evolves complete executable agent safety harnesses co-evolved against adaptive injection stressors, task model/benchmark/tools/evaluator fixed (v1 Nov 17, first publication) - https://arxiv.org/abs/2611.22583
+added:    arXiv 2611.22971 SEAM: 600 multi-step cross-application OSWorld tasks where the instruction and every single screen pass per-screen safety checks but binding a person-fact on one screen to an item-fact on another produces harm - the class CUA benchmarks miss (v1 Nov 17, first publication) - https://arxiv.org/abs/2611.22971
+added:    arXiv 2611.22761: unauthenticated network-adjacent config attacks on the Pollen Robotics Reachy Mini Lite replace robot personality settings and the cloud LLM system prompt through the management interface (CWE-306); 11 prompt-injection attempts across 5 categories were resisted, both config-layer paths changed physical-robot behavior (v1 Nov 17, first publication; companion to the #10-dispositioned bREACHYing-perimeter robot) - https://arxiv.org/abs/2611.22761
+added:    arXiv 2611.21822 vibe-coding threat-landscape systematization: unifies prompt injection, insecure code generation, agent supply-chain attacks and human factors as coding agents execute commands, manage dependencies and touch cloud resources (v1 Nov 16, first publication; same survey rule as #19's deep-research systematization) - https://arxiv.org/abs/2611.21822

-rejected: GitHub global advisories, remainder of the 1,212-row day enumeration (method note below) - unreviewed stream = Oracle November CPU NVD-import dump (00:31:36-00:33:23 bands) + vendor batch 18:31:58-18:32:22 (107 rows per-second recount: F5 BIG-IP token-expiry/APM, Dell PowerStore OS command injection, JetBrains TeamCity x6 build/agent-command RCE classes - TeamCity "agent" = CI agent, keyword noise - PcVue, Azon, WordPress plugins) + steady generic imports; zero AI factor besides the individually dispositioned rows (Pulse naming-only bar)
-rejected: GitHub reviewed-advisory set (15 rows Nov 18) besides the 2 vLLM + LLMKube items: Atomic Server x4 ("token/model" keyword noise, personal file server), Outstatic OAuth proxy, Chainguard SDK SigV4 replay, OliveTin second-order command injection, augment-vir grep() injection = generic platform bugs; install-artifact-from-github CVE-2026-73864 rejected: real PyPI supply-chain fix (unverified native .node artifact via RE2_DOWNLOAD_MIRROR) but not an AI package - spec 2 scopes supply chain to AI packages (msgpack-"mcp" precedent)
-rejected: AI-keyword rows whose mechanism is a generic web bug (AI in product name/output only): AI Puffer WordPress stored XSS CVE-2026-130121 (insufficient escaping of AI-generated comments - sanitization class), Dokan auth bypass CVE-2026-119220, SchemaEngine AI SEO, GamiPress, Nexter Blocks, ZIP AI CSRF (WP plugin classes), IBM Engineering AI Hub x4 (info-disclosure/XSS rows imported 21:31 - mechanism does not touch an AI trust boundary) (#19 AI-SEO-Plugin precedent)
-rejected: MCP advisory wave (same coordinated Nov 18 disclosure batch as the added Obot/Supergateway/Todoist items) held for signal density - real AI-boundary fixes but single-server rows below the individual-item bar: obsidian-sync-mcp CVE-2026-77283 (8.8, no-auth HTTP /mcp default), mssql-mcp-core CVE-2026-63129 (8.8, T-SQL interpolation in MCP DB tools), Gortex CVE-2026-62964 (7.8, absolute-path escape in code-intelligence MCP fileops), Axon CVE-2026-76818 (7.5, unverified prebuilt binary fetch in agent-memory engine), yt-mcp CVE-2026-75612 (7.5, unconfined file_path upload), Network-AI CVE-2026-130750 (6.2, localhost-Origin reflection on McpStreamableServer), NocoDB CVE-2026-130711 (6.5, MCP-token-id role-resolution authz - token is the lookup key, core mechanism is generic authz), DOMShell CVE-2026-75518 (4.8, eval misclassified read-tier), Mattermost CVE-2026-18399 (4.3, GitLab group restriction missing on permalink preview + {{get_projects}} MCP tool) — wave pattern itself is carried by the added items; #5/#19 density rule
-rejected: arXiv Tuesday-batch re-listings with old v1 dates (version history checked per abs page): 2611.21651 (v1 Oct 15), 2511.19218, 2605.07830, 2507.02844, 2507.05248, 2602.07796, 2603.13950, 2609.23466, 2610.16859, 2607.23624, 2607.29241, 2609.01852, 2607.05516 — replacement announcements, not first announcements (#7/#17 age rule)
-rejected: arXiv outside the spec section 2 keyword list: 2611.23022 GPU load/store-unit security (microarchitecture, #19 precedent), 2611.22519 missing security-fix backporting for GitHub Actions (non-ML supply chain), 2611.21777 LLM validators for bridge attacks + 2611.22479 vulnerability-timeline dataset (AI-for-security tooling / non-AI), plus the anonymized-search-dataset, power-grid fault, watermarking (2611.21702, 2611.22345, 2611.23208), forensic-metadata, federated-IDS, CAPEC-model-checking, biometric-policy, mixnet and PIR cs.CR rows + ~300 general agent/ML cs.CL entries matched only on "agent/model/LLM" (#7/#11 keyword rule)
-rejected: arXiv in-niche candidates held out for signal density (abstracts fetched, same bar as #5): 2611.22274 GuardAgent (promotion-injection detection in LLM recommenders), 2611.21804 bias-injection across multi-agent topologies, 2611.22511 LMLineage (context lineage tooling), 2611.22223 GapLeak (privacy leakage in web agents), 2611.22243 shared KV-cache bit flips, 2611.22186 quantization vs safety, 2611.22888 post-stop exploitation budgets on ExploitBench, 2611.22589 LLM honeypots, 2611.22288 ConfGuard monitor, 2611.23074 agent-governance runtime verification, 2611.21641 refusal-aversion measurement — real but secondary to the six accepted
-rejected: OpenAI in-window (7): The Silent Consensus (measures coding-agent acceleration of AI research - metrics post, no disclosure), Research acceleration: jagged frontier (research post), Inside the Codex sandbox (sandbox architecture explainer — hardening/explainer precedent, #19 External-Sandboxes rejection), Managing frontier AI risk program post (program announcement, #10/#11 Daybreak precedent), Figaro partnership, chip-design PPA post (product; Techmeme p45 agrees "up to 14%"), Sumitomo Life customer story — spec 3 (openai.com HTML 403s to curl, dispositioned from the vendor's own RSS summaries, #4 precedent)
-rejected: Anthropic: IPO pricing announcement (funding, spec 3), "Claude solves two open problems in black-hole physics" (research, zero security content), Cyber Verification Program expansion page — datePublished Oct 6 2026 (fetched; outside window) and access-program announcement (#10/#11 precedent); the 17:12:33 sitemap cluster (anthropic-institute-agenda, 81k-economics, AI-assistance-coding-skills, Sonnet-3.7 economic index, evaluating-ai-systems, coding-agents-social-sciences, companionship, economic-index — identical-second template re-edits of old posts) = site-wide churn, no new post (#12/#13/#18 precedent)
-rejected: Microsoft: Security Blog feed latest = Nov 17 Perception post (dispositioned #20), zero in-window; MSR blog feed answered on the 4th try after three 504s, latest still the Nov 17 health-AI post (#20); microsoft.ai "Seven new world-class models at scalable price-performance" (Ignite model launch, HN 415, fetched 200) — product/marketing bar
-rejected: AWS in-window: "Step-up authentication: right-size authentication to risk" + "Detecting cross-cloud lateral movement with AWS Security Hub" — product how-to posts, zero disclosed vulnerability or novel attack technique, zero AI factor (#10/#11/#13 hardening precedent)
-rejected: Meta newsroom in-window: Commercial Terms of Service update + Loblaw retail partnership — no security content
-rejected: BleepingComputer in-window remainder: Ukraine/FBI ransomware arrests, Coinbase blocks $4.7M Musk-deepfake transfers (fraud-prevention news, no primary artifact), 40 Irish credit unions IT-provider breach (zero AI factor), "From Deepfakes to Helpdesk Impersonation" verification post, Dizigom takedown, Apache Roller RCE exploited in attacks (classic CMS RCE, zero AI factor), Windows Server 2028 announcement, Scattered Spider sentencing (judicial) — spec 3
-rejected: GitHub changelog (4): SaaS connectors for Copilot, Gemini 4 Neon availability, "Security risk assessments run automatically on GitHub Team" + "Customize Copilot for your team" — model availability + governance/feature launches, not disclosures (#10-#19 governance-feature precedent)
-rejected: coding-agent changelogs — Codex rust-v0.170.0 stable (Nov 18 03:40, full body fetched: WebRTC voice + OTel auth-header features, zero security-fix mentions) + single-line alphas (#6-#20 handling); Copilot CLI v1.0.114 body = "Fixes and changes"; Claude Code nothing newer than v2.1.327 (#20); Cursor changelog newest entry still Nov 6
-rejected: Hacker News >=150 band, 20 stories — besides the added NGINX RCE (1069): Gemini 4 Neon (338) + Genie 4 (169) + Microsoft seven models (415, above), SEO/GEO traffic study (689), DNS-piracy essay (570), Inside Deutsche Bank (525), ID.me IRS terms (491), YC interview story (486), Russia/Poland report (341), KPMG audit pilot coverage (320), Energy-Based Reasoners (297), Codeberg outage (295), Gemini Carbon complaint (262), dictionary/misc items (228, 228, 201), KVM CVE-2026-99312 (165, kernel bug, zero AI factor), "Claude was trained on protein sequences" (160, training-data provenance research, no security disclosure) — zero further in-scope disclosures with artifacts
-rejected: Hacker News keyword band, notable rows: FBI-ARTEX coverage re-surfacing (Industrial Cyber + The Next Web, both 403 to the validator; FLASH-20261113-001 was added in #15 and FBI's own Nov 17 post says no additional warning was issued — repeat coverage, #16 precedent); "AI Repo Man: Exploiting Agentic AI Debt Collection Platforms" (eaton-works.com) — source URL 404s to the validator with no second source to corroborate (spec 3: reject with reason); Bitrise preliminary incident review re-surfacing (#13 class, zero AI factor); FT/Washington Sun/TechPolicy coverage of bank-CEO + NSA congressional testimony blaming AI-generated patches for the US outages (outage-aftermath reporting, no primary artifact, #13 outage precedent); EFF "AI Cybersecurity Policy Framework" (policy advocacy, not a disclosure); cryptologie.net "The Wong test" (named researcher but a workflow essay, no PoC/artifact, spec 3); Ruby/WordPress 7.1.5/PHP/Shiro security releases (non-AI platform patches); Airbyte + Bunny.net security-incident status posts (zero AI factor); Krebson Optimizely A/B-test-malware coverage (= OPTIMJACK, dispositioned #20); CISA Java4Shell emergency alert (= item added #20, follow-up coverage); South Korea AI-cheating bounties (policy); Gallup companionship survey; Show HN/product launches (Intrusive.ai, AgentShield, KradleBench, Otari, Space-Core...) — spec 3
-rejected: Techmeme in-window (15 items): DHS Threat Disruption Center launch (program announcement, zero disclosure, #10/#11 precedent), Palo Alto Q1 revenue + Applied Compute acquisition, Sierra raise, Remanence seed, Meta Class-C pause, Nvidia all-hands, Gemini Tasks/Genie 4/Neon product rollouts, DNI/DHS/FBI "no outage-attack link" testimony coverage (non-AI outage aftermath), OpenAI chip PPA post (= OpenAI rejection above), Newsom AI-regulation proposal (policy proposal — unlike the #18 California suit there is no filed-complaint artifact), Anthropic IPO pricing — spec 3
-rejected: Hugging Face blog 5 in-window posts (model releases + inference-engineering posts, zero security content); Unit 42 (latest Nov 17 OPTIMJACK, dispositioned #20); Google Project Zero (latest Nov 13, #12); Embrace The Red (latest Oct 13); GitHub blog (latest Nov 17, #20); JFrog (Java4Shell + AccuraSoftHire posts dispositioned earlier; 2 posts added above); ReversingLabs other Nov 18 lastmods (AWS-account landing page + Agentic SOC event page — marketing); Phrack (no new article since #20's Issue-73 disposition except the backdoor paper added above; no Issue 74)

Swept, zero new items in window: OpenAI news RSS (7 in-window posts dispositioned above), Anthropic
sitemap (churn cluster + 3 new pages dispositioned), Google Project Zero (feed.xml latest Nov 13),
Microsoft Security Blog + Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red,
BleepingComputer (15-item feed, 9 in-window dispositioned), arXiv both channels (Tuesday batch
fully dispositioned), GitHub blog, Hugging Face blog, JFrog Research, ReversingLabs, per-repo advisory
listings llama.cpp (newest Mar 26) / vLLM (newest Nov 18, 2 added above) / transformers (none) /
LangChain (Jun 12) / CrewAI (zero) / PyTorch (Nov 10, added #6), NVD keyword searches Nov 18 00:00 ->
Nov 19 00:00: jailbreak 0, "prompt injection" 1 (= yt-mcp, held in the wave above), LLM 2 (= LLMKube
added + yt-mcp), "AI agent" 5 (all in the wave above), MCP 17 (all individually dispositioned above).
Daily Swig: connection still fails from this host (curl exits without an HTTP response), Q1 stays
open. blog.pypi.org answered 429 again (#14/#15 precedent); no candidate depended on it this round.

Method notes (not policy questions):
- GitHub advisories window: the page parameter is ignored when a published= range is set (asc pages
  1-8 byte-identical), so enumeration = descending page + ascending frontier walks (each slice <=100
  rows or boundary-adjacent) + per-second recount of the two batch bands: 00:31:36-00:31:51 sums to
  exactly 100 rows with zero saturated seconds (the Oracle dump publishes ~7/s), 18:31:58-18:32:22
  recounts 107 rows, tail union for 18:32:23->24:00 = 103 rows with zero rows after 21:55:09
  => 1,212 unique rows, all boundaries contiguous, full-description keyword scan (159 raw hits, 22
  strong-AI-word hits, all dispositioned individually above).
- NVD: pubStartDate/pubEndDate range queries now 404 on every combination including date-only
  (API-side breakage; the #13-#20 window-query method is dead). Fallback used: full keyword sets
  fetched (jailbreak 15, "prompt injection" 226, LLM 373, "AI agent" 343, MCP 932 rows, all complete)
  with the Nov 18 published window filtered client-side. Because the unreviewed GHSA stream mirrors
  NVD, this scan also covers the Oracle/vendor import dumps; nothing with an AI keyword outside the
  rows dispositioned above exists in the window.
- HN window: created_at_i 1794960000-1795046400 (Nov 18 00:00 -> Nov 19 00:00 UTC), 1,263 unique
  stories fetched across three adjacent sub-windows (Algolia page cap 1,000).
- Invisible-dots browser was NOT needed: every source and every added-item URL answered plain curl
  with 200 (openai.com dispositioned from its RSS per #4). socket.dev, industrialcyber.co,
  thenextweb.com and eaton-works.com blocked or dropped the validator, and none of them was the
  primary source of any accepted item; the eaton-works write-up was rejected for exactly that reason.
  Per operator instruction all invisible-mcp-vnc containers are stopped and removed after the sweep.

## 2026-11-19 12:00 UTC — sweep #22

Window: Nov 19 00:00 UTC -> Nov 19 12:00 UTC (12h). Sources swept: OpenAI news (RSS), Anthropic
(sitemap lastmod, re-checked in-window), Google Project Zero (feed.xml, new URL — feed.xml answers,
feeds/posts/default 404s), Microsoft Security Blog + Microsoft Research blog, Meta newsroom, AWS
Security Blog, Unit 42, Embrace The Red, Daily Swig (reachability retest), BleepingComputer, arXiv
cs.CR+cs.CL (Wednesday batch landed 01:00 UTC), GitHub global advisories API (published-range
boundary walk: 100 unique rows, next slice below boundary returned 0 rows = window complete;
full-description keyword scan 28 raw hits, 4 dispositioned individually), per-repo security-advisories
API (llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch), NVD keyword searches (jailbreak /
prompt injection / LLM / "AI agent" / MCP, client-side Nov 19 filter), Hacker News (all 386 in-window
stories: >=150 band + keyword scan of 95 hits), Techmeme (feed.xml, 15 items), Hugging Face blog,
ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog + changelog, coding-agent
changelogs (Claude Code, Codex, Copilot CLI, Cursor), Phrack issue list, microsoft.ai news.
All 4 added URLs fetched this session, HTTP 200, paper title/ID tokens verified present in the
fetched body; every individually rejected candidate was also fetched or feed-verified this session.

+added:    arXiv 2611.24231 measures which prompt-injection operations survive rendering untrusted documents as page images: pairing text and page-image versions of 191 QA sources across six models, rendering lowers append/format execution for every model, yet Qwen3-VL-8B still executes image-borne answer-replacement at 77.96%; failed image attacks are often commands read but declined (models transcribe them), trusted-user authorization raises execution on unchanged pages, and an 8-candidate attack search largely closes the image deficit for Qwen and Gemini (v1 Nov 18, first publication) - https://arxiv.org/abs/2611.24231 (200, title + submitted Nov 18 in body)
+added:    arXiv 2611.24313 AutoPI: taxonomy-guided adaptive red teaming for prompt injection in LLM agents — attacker LLM writes one injection per PIAtlas sub-technique and reworks failures from judge feedback; on AgentDojo finds a successful injection against Qwen3.6-35B in all six taxonomy classes and >90% of techniques, 20-50% higher ASR and 25-400% higher class coverage than recent adaptive red teamers against Claude Haiku 5.5; code + attacks released (v1 Nov 18, first publication) - https://arxiv.org/abs/2611.24313 (200, AutoPI + PIAtlas tokens in body)
+added:    arXiv 2611.24362 PIPETTE: shows red-teaming ASR does not translate into adversarial-training robustness — RL-Hammer and PISmith mode-collapse into repeated attack strategies that add zero robustness; purple-teaming framework with dictionary-learning coverage reward + on-policy adversarial training reaches near-zero ASR on static attacks and holds under adaptive attacks at near-baseline utility (v1 Nov 18, first publication) - https://arxiv.org/abs/2611.24362 (200, PIPETTE + RL-Hammer tokens in body)
+added:    arXiv 2611.24076 closed-loop steering of the released GR00T N1.7 VLA robot policy: landmark synonyms absent from training instructions send the policy to the bowl they name (67.6% vs 100% completion), and replacing only the landmark tokens' vectors before the VL-to-action connector sends nearly every episode to the other bowl; connector-only finetuning on diversified landmark words suffices for the repair (v1 Nov 18, first publication; cross-list announcement — v1 submitted Nov 18, so not a re-listing) - https://arxiv.org/abs/2611.24076 (200, GR00T + connector text in body)

-rejected: GitHub global advisories, remainder of the 100-row window (NVD-import batches 00:30/03:31/06:30/09:30) — Vtiger CRM x5, eladmin x2, Geany build-command injection, yt HDF5 con_args eval (scientific-viz package, not AI), SourceCodester x4, SEO Panel x2, WordPress plugins x3, cassandra-jdbc-wrapper, PGPainless, Hyperev, Idunno, Unblu, RAGit-not-AI rows: zero AI factor besides the individually dispositioned rows (Pulse naming-only bar); note the sort=published&direction=desc params (sort=desc alone now 422s)
-rejected: mssql-mcp-core CVE-2026-63129 (8.8) — same MCP wave #21 held for signal density and dispositioned there; GitHub mirror published it Nov 19 00:46, no new facts; Q12 open, density rule held per its own 2-sweep default (Q12 pending)
-rejected: MemTensor memmy-agent CVE-2026-132560 (3.3, local-only information disclosure of /api/v1/memory/logs, "unknown function" CNA boilerplate) — AI product name only, mechanism does not touch an AI trust boundary (#19 AI-keyword-naming precedent)
-rejected: arXiv outside the spec section 2 keyword list: 2611.24058 MIRAGE (Knowledge Injection = knowledge-editing task, not an attack), 2611.23701 LDP triangle-counting poisoning (privacy), 2611.24623 LDP histogram dropout, 2611.24662 CPS system-identification poisoning defense (control systems), 2611.23660 + 2611.23669 web-attack detection (AI-for-security tooling), 2611.24489 + 2611.23304 + 2611.23267 vulnerability detection/prediction (AI-for-security tooling, #21 2611.21777 precedent), 2611.24765 LLM-guided smart-grid attack paths (AI-for-security), 2611.23764 IoT Ethereum auth, 2611.23826 5G slice isolation, 2611.23666 multi-tenant DC workload-sync attacks (zero AI factor) + ~160 general cs.CR/cs.CL entries matched only on agent/model keywords (#7/#11 keyword rule)
-rejected: arXiv re-listings with old v1 (version checked per abs page or ID band): 2407.10445 I2I Backdoor (2024, cross-listed today), 2610.38786 (v1 Oct), 2611.19763 (v1 Oct ID band) — replacement announcements, not first announcements (#7/#17 age rule); HN-surfaced: 2607.29378 previous-token LLM inversion (v1 Jul), 2610.40716 module-shadowing agent study (v1 Oct)
-rejected: arXiv in-niche candidates held out for signal density (abstracts fetched, same bar as #5/#21): 2611.23770 DetectViT (test-time backdoor detector for ViTs — detector class), 2611.24494 RECLAIM (training-free VLA patch defense), 2611.24558 SCR (poisoned-sample detector, CIFAR-scale), 2611.24936 JDPI (jailbreak detector vs Llama Guard), 2611.24682 state-dependent command-safety answerability audit (15-pair pilot) — real but secondary to the four accepted; #5/#19/#21 density rule
-rejected: BleepingComputer in-window: Edge Lockdown Mode GPU-API test (feature preview, zero disclosure), Windows hotpatch app-launch issues (IT operations news), NVIDIA investigating Gentlemen ransomware leak-site claim (corporate breach claim, zero AI factor — Bitrise/Irish-credit-unions precedent)
-rejected: Anthropic sitemap: research/formal-verification lastmod Nov 18 13:50 fetched — datePublished Oct 22 (LLM formal-verification research post, zero security disclosure; edit, not a new post; #12/#18 churn rule); zero Nov 19 lastmods in-window; OpenAI, MSR, Meta newsroom, AWS, Unit 42, GitHub blog, HF blog, ReversingLabs, JFrog latest entries all <= Nov 18 and dispositioned in #21
-rejected: Hacker News >=150 band, 4 stories besides zero added: smoltcp 15M-web-server VM (networking demo), Ruby2ßhell pre-auth Rails RCE (181, classic web-stack bug, zero AI factor), Gemini 4 Neon vs Opus 5.5 benchmark comparison (model bake-off, no disclosure), UT2004 in browser — spec 3
-rejected: Hacker News keyword band, notable rows: PyPI agent-framework coverage re-surfacing (The Register, BleepingComputer, OpenSourceMalware — item added #21, follow-up coverage rule #16), The Register NGINX AI-discovery coverage (= item added #21), Phrack 73 backdoor paper re-surfacing (= item added #21), Rails trilogy SQLi + sqlite3-ruby truncation advisories (non-AI platform), TfL photocard breach 265k (zero AI factor), Optimizely initial-incident PDF (= OPTIMJACK, dispositioned #20/#21), Meta Muse agent safety-mishaps Reuters investigation (product-safety journalism, anonymous sourcing, no primary artifact/PoC), Newsom AI kill-switch proposal (policy proposal, no filed artifact — #21 Newsom precedent), Anthropic IPO coverage wave x8 (funding, spec 3), arXiv 2611.23174 inference-after-coding-agents (selective-inference statistics position paper, zero security disclosure), zimmergren.net AI-security-engineering essay (opinion, no PoC/artifact, spec 3), Claude Code subagent duplicate-report gist (functional bug, no security impact), Vercel Eve GA / Frostbite AI-designed CPU / Cloudflare Clairic weights (product launches), Doritos-gun lawsuit coverage (AI harm, not security) — spec 3
-rejected: Techmeme in-window (15 items): TikTok EU local-owner proposal, Android foldable mandate, Snap Easy Lens CLI, "an argument against the silent consensus" (op-ed on OpenAI's already-dispositioned post), Bezos Prometheus $18B raise, Anthropic PBC listing + UK/EU retail-sale filings (funding), Alibaba Q2 revenue, Moonshot AI contract cancellation, South Korea export record, TSMC/Samsung China shipment pause (non-AI chip geopolitics), Netflix CCO profile, campus AI-mandate pushback, robot order stats, Porat grid-investment remarks — spec 3, zero disclosures
-rejected: GitHub changelog in-window: S3 audit log streaming OIDC on GHE.com + wait-for-PR-events GitHub CLI — platform features, zero disclosure (#10-#21 governance-feature precedent); Copilot SaaS-connectors item dispositioned in #21
-rejected: coding-agent changelogs — Claude Code v2.1.328 (Nov 19 01:14, full 19.2k-char body fetched: auto mode GA on Bedrock/GCP/Foundry, mcp_tool hook onFailure:block, self-hosted-runner features; sandbox lines are a .git/HEAD rename bug, a hook-rewrite behavior change and failIfUnavailable reporting — zero sandbox-escape/CVE/security-fix lines, unlike the #4 worktree-escape line); Codex nothing newer than Nov 18 releases (dispositioned #21); Copilot CLI nothing newer than v1.0.114 (#21); Cursor changelog newest entry still Nov 6
-rejected: Daily Swig connection still fails from this host (curl exits without an HTTP response), Q1 stays open; blog.pypi.org answered 429 again (#14/#15/#21 precedent), no candidate depended on it; microsoft.ai feed answers an HTML shell, zero entries (#21); Phrack latest issue still 73, no Issue 74

Swept, zero new items in window: OpenAI news RSS (latest Nov 18 10:00, all #21), Anthropic sitemap
(zero Nov 19 lastmods), Google Project Zero (feed.xml latest Nov 13), Microsoft Security Blog +
Research blog (latest Nov 17), Meta newsroom (Forum product post Nov 19, zero security), AWS Security
Blog (2 Nov 18 posts dispositioned #21), Unit 42 (Nov 17), Embrace The Red (Oct 13), Hugging Face
blog (latest Nov 18 19:54, model/engineering posts), GitHub blog (Nov 17), ReversingLabs (Nov 18 =
#21), JFrog Research (Nov 18 = #21), per-repo advisory listings llama.cpp (newest Mar 26) / vLLM
(newest Nov 18, 2 added #21) / transformers (none) / LangChain (Jun 12) / CrewAI (zero) / PyTorch
(Nov 10, added #6), NVD keyword searches Nov 19 00:00 -> 12:00: jailbreak 0, "prompt injection" 0,
LLM 0, "AI agent" 0, MCP 0.

Method notes (not policy questions):
- GitHub advisories window: published-range boundary walk (sort=published&direction=desc; plain
  sort=desc now 422s); page-1 tail 00:30:34 and the sub-boundary slice 00:00-00:30:33 returned 0
  rows, so the 100-row enumeration is contiguous and complete; per-second batches at 00:30:3x /
  03:31:3x / 06:30:3x / 09:30:3x are the NVD import mirrors, consistent with yesterday's shape.
- NVD: pubStartDate/pubEndDate range queries still broken; fallback used again — full keyword sets
  fetched (jailbreak 15, prompt injection 226, LLM 373, AI agent 343, MCP 932) with client-side
  Nov 19 filter; result all-zero agrees with the GitHub unreviewed stream above (Q13 open).
- HN window: created_at_i 1795046400-1795089600 (Nov 19 00:00 -> 12:00 UTC), 386 stories single page.
- arXiv: Wednesday batch (announced Nov 19 01:00 UTC) = first distribution for 2611.236xx+ IDs;
  #21's enumeration stopped at 2611.23208, no overlap. All 4 accepted papers verified v1 = Nov 18
  on their abs pages (24076 is a cross-list announcement but its v1 is Nov 18, so not a re-listing).
- Invisible-dots browser was NOT needed: every source and both added-item URL families answered plain
  curl with 200 (openai.com dispositioned from its RSS per #4; bleepingcomputer.com, unit42, techmeme,
  microsoft, huggingface, arxiv, github APIs all clean). Per operator instruction all
  invisible-mcp-vnc containers are stopped and removed after the sweep.
## 2026-11-20 00:00 UTC — sweep #23

Window: Nov 19 12:00 UTC -> Nov 20 00:00 UTC (12h). Sources swept: OpenAI news (RSS), Anthropic
(sitemap lastmod), Google Project Zero (feed.xml), Microsoft Security Blog + Microsoft Research
blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red, Daily Swig (reachability
retest), BleepingComputer, arXiv cs.CR+cs.CL (both RSS channels; no new announcement batch in
window), GitHub global advisories API (published= range walk: 887 unique in-window rows,
contiguous slices, full-description keyword scan), per-repo security-advisories API (llama.cpp,
vLLM, transformers, LangChain, CrewAI, PyTorch), NVD keyword searches (jailbreak / prompt
injection / LLM / "AI agent" / MCP, client-side window filter), Hacker News (all 779 in-window
stories: >=150 band + keyword scan), Techmeme (feed.xml, 15 items), Hugging Face blog,
ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog + changelog, coding-agent
changelogs (Claude Code, Codex, Copilot CLI, Cursor), Phrack issue list, microsoft.ai news. All 4
added URLs fetched this session, HTTP 200, with the product/CVE/GHSA/paper tokens verified
present in the fetched body; every individually rejected candidate was also fetched or
feed-verified this session.

+added:    CrewAI GHSA-wp6h-6xxh-g8fp (high, watchlist repo, Nov 19 18:05): a security review of crewAI 1.15.30 found three bypasses in previously hardened components - _resolve_tool_dict() resolves dotted-path cache_function callables and imports arbitrary tool_type modules even when the CREWAI_DESERIALIZE_CALLBACKS guard is disabled (the guard exists precisely to stop untrusted checkpoints/serialized payloads importing arbitrary modules), NL2SQLTool allow_dml=False read-only validation can be bypassed, and FilePath._validate_file_exists accepts symlink sibling-prefix and parent-symlink paths; crewai, crewai-tools and crewai-files fixed in 1.15.31 - https://github.com/crewAIInc/crewAI/security/advisories/GHSA-wp6h-6xxh-g8fp (200, CREWAI_DESERIALIZE_CALLBACKS + NL2SQLTool in body)
+added:    Model Context Protocol Python SDK GHSA-qx49-fqc8-xw99 / CVE-2026-115563 (high, Nov 19 22:45): the SDK's OAuth client support (mcp.client.auth) let the MCP server a client is connected to decide where the client's OAuth credentials go - authorization-server metadata issuer not validated on every discovery path and stored/pre-provisioned credentials not bound to their server - so a malicious or compromised MCP server (naming its own authorization server in protected-resource metadata, or publishing none and presenting the user's real server as issuer) receives the client_secret, authorization code and PKCE code_verifier - or a signed client assertion - meant for the real server - https://github.com/modelcontextprotocol/python-sdk/security/advisories/GHSA-qx49-fqc8-xw99 (200, mcp.client.auth + code_verifier in body)
+added:    ServiceNow centrally remediates an indirect prompt injection issue in the Email Response AI Agent of Now Assist for CSM (CVE-2026-18883, high; Nov 19 vendor batch, NVD 20:17): deployed centrally with no version upgrade, hot fix or backport required, no known exploitation; the same vendor batch also closed a code injection issue in the ServiceNow AI Platform (CVE-2026-125288) - https://github.com/advisories/GHSA-f7fm-26fm-r3gq (200, "indirect prompt injection" + "Email Response AI Agent" in body)
+added:    arXiv 2611.23346 state-transfer auditing of long-horizon LLM agents: critical security failures sit in the harness's state transfers, not the prompt/tool boundary - transfers admit unchecked content (indirect prompt injections, unverified memory writes), rewrite state lossily (compaction dropping user constraints, consolidation stripping source authority) and restore incompletely (rollbacks leaving aborted state); proposes state-transfer auditing and harness controls (v1 Nov 17, Wednesday batch first publication; missed by #22's per-item scan - this sweep's channel-vs-log cross-check caught the gap, dispositioned here) - https://arxiv.org/abs/2611.23346 (200, title + ID tokens in body)

-rejected: Splunk AI Assistant x4 (CVE-2026-131302 medium, 131303/131304/131305 high, same Nov 19 21:31 vendor import as ServiceNow): users without the admin/power Splunk roles could submit client-controlled feature settings through the REST API and modify limited response metadata outside the server-side configuration boundary - missing role/config enforcement in an AI product, the mechanism does not touch a model trust boundary (#19 AI-keyword-naming bar)
-rejected: MCP single-server wave, held for signal density (same logic as #21/#22; the wave pattern is carried by the added MCP Python SDK item): k8s-mcp-server CVE-2026-130007 (high, missing authentication on default SSE/HTTP transport), sqlite-mcp CVE-2026-130008 (high, read-only bypass via statement chaining + SQL injection in describe_table), Safari MCP GHSA-4rj4-rcrq-gr92 (medium, JS injection through single-quoted tool arguments, no CVE), Jupyter MCP Server CVE-2026-77318 + CVE-2026-77359 (pre-1.0.3 standalone streamable-http mode; NVD 17:18, never individually surfaced before - single-server rows), Flyto2 Core CVE-2026-77314 (pre-2.26.8 MCP run_recipe tool; NVD 22:18) - #21 wave-density precedent
-rejected: CodeWhale (DeepSeek-TUI coding agent) x2 (CVE-2026-83816 + CVE-2026-83817, high, maintainer-validated, fixed 0.9.0 Nov 19 22:48): the exec_shell "parallel read-only" classifier auto-approves fd/rg/git grep exec flags enabling silent RCE despite the documented every-shell-command approval prompt, and Automation auto_approve defaults to true bypassing the CVE-2026-45374 task-path hardening - a genuine agent approval-gate class held for signal density against this sweep's three advisories; flag for reconsideration as a standalone item if the coding-agent approval-gate genre keeps moving (#21 wave-hold precedent)
-rejected: Pydantic AI Harness CVE-2026-133161 (high, fixed 0.70.0): strip_ansi=True ANSI-escape pattern in ToolOutputLimits is quadratic on crafted unterminated sequences - event-loop DoS through a sanitizer regex, no trust-boundary crossing
-rejected: Guard Core CVE-2026-129302 (high): quadratic ReDoS in built-in penetration-detection patterns (regression since 3.5.0 - built-ins bypass the timeout-protected matcher) - web guard middleware, zero AI factor (#21 naming-only precedent)
-rejected: NLTK x3 (CVE-2026-128138 critical, CVE-2026-129234 + CVE-2026-129221): additional JVM/JAR sandbox bypass through unvalidated launcher tokens in java(), untrusted search path in prover9/mace4 config, Weka ARFF schema-directive injection via class labels - vulnerable surfaces are the Weka/Prolog launcher bridges: classic library bugs in an NLP library, not an AI model/agent trust boundary (#21 mechanism-generic precedent)
-rejected: Mistune CVE-2026-97698 (high): quadratic parse_link_href scan on empty-label links - markdown parser DoS, zero AI factor
-rejected: Nov 18 coordinated MCP disclosure batch re-published into the global advisory DB Nov 19 20:34-20:41: DOMShell CVE-2026-75518, Axon CVE-2026-76818, Gortex CVE-2026-62964, obsidian-sync-mcp CVE-2026-77283 (all individually held for density in #21) + Todoist CVE-2026-76807 (= item added #21) - re-publish with zero new facts; #16 follow-up coverage rule, #22 mssql-mcp-core precedent
-rejected: arXiv 2611.23844 envelope-layer defense measurements (v1 Nov 18, #22 batch, not individually named in the log - dispositioned by this sweep's channel-vs-log cross-check): measures component weight/arbitration when a pre-planted malicious MCP tool and poisoned long-term memory coexist in a single agent - real but a measurement/defense-component paper secondary to this sweep's adds (#22 GuardViT/JDPI hold bar)
-rejected: arXiv 2611.24689 exact training-data stealing from other federated-learning clients (same cross-check catch, dispositioned here): FL privacy-extraction genre - no AI product trust boundary touched (#21 GapLeak hold precedent)
-rejected: arXiv current-channel keyword noise (feed-verified, title-level): 2611.23274 pre-registered conceptual replication study, 2611.23483 SpecLook agentic-MLLM inference speedup, 2611.23570 agent message-skip latency tradeoff - zero security disclosure (#7/#11 keyword rule)
-rejected: arXiv current-channel cross-listings with old v1 (2608.25697 LLM-interpretability runtime defense, 2609.28559 coding-agent model-identity fingerprinting, 2609.32495 agent-harness audit-record audit, 2605.24173 diffusion-LM training-data extraction, 2606.23496 TROPT text-optimization framework, 2607.01919 ElephantAgent state continuity, 2503.01935 MultiAgentBench, 2602.15198 Colosseum collusion auditing): replacement announcements, not first announcements (#7/#17 age rule)
-rejected: GitHub global advisories remainder of the 887-row window: unreviewed stream = NVD import dumps (12:31:47-12:32:03 and 15:31:15-15:31:27 bands, also covered client-side through the NVD keyword fetches) + the 21:31:41-21:32:07 vendor batch (189 rows, max 15 rows/sec, zero per-second saturation; the ServiceNow + Splunk AI rows dispositioned above) + steady generic imports; zero AI factor beyond the individually dispositioned rows (Pulse naming-only bar)
-rejected: OpenAI in-window zero (RSS latest Nov 19 08:00 Performance Factory metrics post, outside the window; the two Codex Security Plugin / PR-scans demo pages stamped 00:00 are pre-window and marketing pages anyway) — spec 3
-rejected: Meta in-window: India government-led scam-awareness microdrama premiere (PSA/marketing) - spec 3
-rejected: AWS in-window 3: re:Invent security guide (event promo), WAF security analytics dashboard (product feature), Security Incident Response triage how-to - zero disclosure, zero AI factor (#10-#22 hardening/how-to precedent)
-rejected: GitHub blog "The best things in open source are built together" (community post) + GitHub changelog 2 (Azure DevOps Enterprise Live Migrations GA, RHEL runner images GA) - platform features, zero disclosure (#10-#22 governance-feature precedent)
-rejected: Anthropic sitemap: homepage lastmod 15:06 (template churn) + /research/formal-verification re-edited 13:21 = the same datePublished Oct 22 post dispositioned in #22 - edit, not a new post (#12/#18 churn rule); zero new Nov 19 afternoon posts
-rejected: ReversingLabs in-window 4 lastmods: two marketing pages (SentinelOne investment, scalable file analysis), the Java4Shell "is inside the software you run" post (Java4Shell dispositioned #20; lastmod edit, no new facts) + Turla Cereal 12 Exchange loader hunt post - APT loader telemetry with zero AI factor - spec 3
-rejected: BleepingComputer in-window 7: SBC Medical Group 1.6M-customer breach (healthcare breach, zero AI factor), ASUS Control Center Express hard-coded key (non-AI platform), TfL Oyster 265k breach (= #22-rejected, re-surfacing), PHP 18-flaw security update (non-AI platform), ransomware warning-signs explainer (advice content), NGINX2Root coverage (= item added #21, follow-up coverage #16), Drupal critical RCE patch (non-AI platform) - spec 3
-rejected: Hacker News >=150 band, 10 stories: Anthropic Nasdaq debut 828 (funding, spec 3), pancreatic cancer halted by virus injection (medicine), UK government secret Apple encryption order (policy, zero AI factor), "Fuck Your AI Safeguards" (opinion essay, no PoC/artifact, spec 3), Claude black-hole physics 346 (= post dispositioned #21, re-surfacing), survival book site, foldable-zine app, Pete Warden hardware podcast, paid-advertising essay - zero disclosures with artifacts
-rejected: Hacker News keyword band, notable rows: Arizona investigating OpenAI + other firms over "possible rogue agents" (policy investigation, no filed artifact/complaint, #21 Newsom precedent), Synology SA_26_21 advisory (non-AI NAS platform patch), "OpenAI still has a WebRTC Problem" (networking engineering post, zero disclosure), HubSpot Aviator agent harness (engineering-blog product post, no disclosure), "An FBI agent called a Claude user about his chats" (policy essay, no artifact), Chrome AI features wasting bandwidth (product complaint), Doritos-gun lawsuit coverage (AI harm, not security), Anthropic IPO coverage x3, kernel/PHP/Synology security-release rows (non-AI platforms) — spec 3
-rejected: Techmeme in-window (15 items): Nvidia-Inception acquisition talks, National Cyber Director departure, Apple coding-agent job listing (product), Intuit selling Mailchimp (funding), CISA director-nominee hearing (policy), Meta Muse Charm pricing (product), AI Safety and Innovation Caucus (policy), Gemini macOS computer use (product rollout), Everful + Boltz raises + Alphabet CA$5B bond (funding), Kalshi self-exclusion lists, Ubisoft earnings, Google "Mirror Dimension" isolated-internet testbed for browsing agents (evaluation-infrastructure reporting, zero disclosure), Anthropic IPO close - spec 3
-rejected: coding-agent changelogs - Claude Code v2.1.329 (Nov 19 22:14, full 18.1k-char body fetched: the two sandbox mentions are a /sandbox allowUnixSockets display fix and claude doctor Keychain advice on macOS - zero sandbox-escape/CVE/security-fix lines, unlike #22's analysis of v2.1.328); Codex rust-v0.172.0-alpha.2 + v0.171.0-alpha.13.1 single-line alphas (#21 handling); Copilot CLI v1.0.115-0 "Fixes and changes" + v1.0.115-1 (MCP sign-in refresh-scope continuation + workflow-register timing = functional fixes, no security disclosure); Cursor changelog newest entry still Nov 6
-rejected: Daily Swig connection still fails from this host (curl exits without an HTTP response), Q1 stays open; blog.pypi.org answered 429 again (#14/#15/#21/#22 precedent), no candidate depended on it; microsoft.ai feed answers an HTML shell, zero entries (#21/#22); Phrack latest issue still 73, no Issue 74

Swept, zero new items in window: OpenAI news RSS (latest Nov 19 08:00, outside window), Google
Project Zero (feed.xml latest Nov 13), Microsoft Security Blog + Research blog (latest Nov 17),
Unit 42 (Nov 17 OPTIMJACK wrap-up, #20/#21), Embrace The Red (Oct 13), Hugging Face blog (latest
Nov 18 19:54), JFrog Research (Nov 18 = #21), GitHub blog (Nov 17), per-repo advisory listings
llama.cpp (newest Mar 26) / vLLM (newest Nov 18, 2 added #21) / transformers (none) / LangChain
(Jun 12) / CrewAI (Nov 19, 1 added above) / PyTorch (Nov 10, added #6), NVD keyword searches
Nov 19 12:00 -> Nov 20 00:00: jailbreak 0, "prompt injection" 2 (= added ServiceNow item), LLM 4
(Linux kernel import noise), "AI agent" 2 (ServiceNow + Pydantic AI, dispositioned above), MCP 6
(Pydantic AI + HID mcp2200/mcp2221 kernel noise + Jupyter x2 + Flyto2, dispositioned above).

Method notes (not policy questions):
- GitHub advisories window: the published_start/published_end parameter pair now silently ignores
  published_end (eight walks with different ends returned the identical newest 100 rows); the
  documented published=START..END range is honored on both boundaries, so enumeration = recursive
  range splits on 100-row saturation: 11 slices, all boundaries contiguous and every resolved
  slice <100 rows => 887 unique rows, complete; the hh:31-:32 import bands mirror the NVD
  12:17/15:17 publishes and are doubly covered by the client-side NVD keyword fetches.
- NVD: date-range queries still broken (#13-#22); full keyword sets fetched (jailbreak 15,
  "prompt injection" 228, LLM 377, "AI agent" 345, MCP 938 rows, all complete) with the
  Nov 19 12:00 -> Nov 20 00:00 window filtered client-side.
- HN window: created_at_i 1795089600-1795132800 (Nov 19 12:00 -> Nov 20 00:00 UTC), 779 stories
  single page.
- arXiv: no new announcement batch inside the window - the Thursday batch announces Nov 20 01:00
  UTC, an hour after the window closes, so it belongs to the next sweep. The live RSS channel is
  the batch #22 dispositioned; a systematic channel-entries-vs-sweep-log cross-check (new this
  round) found the 2611.23xxx-235xxx ID band under-individualized in #22's bulk keyword rejects:
  2611.23346 added above, 2611.23844 + 2611.24689 individually held above, remaining band entries
  re-verified keyword-noise or old-v1. The cross-check is now part of the standard sweep method.
- Invisible-dots browser was NOT needed: every source and all four added-item URLs answered plain
  curl with 200 (openai.com dispositioned from its RSS per #4; bleepingcomputer.com, unit42,
  techmeme, microsoft, huggingface, about.fb.com, arxiv and the GitHub APIs all clean). Per
  operator instruction all invisible-mcp-vnc containers are stopped and removed after the sweep.

## 2026-11-20 12:00 UTC — sweep #24

Window: Nov 20 00:00 UTC -> Nov 20 12:00 UTC (12h). Sources swept: OpenAI news (RSS), Anthropic
(sitemap lastmod), Google Project Zero (feed.xml Atom entries), Microsoft Security Blog + Microsoft
Research blog, Meta newsroom, AWS Security Blog, Unit 42, Embrace The Red, Daily Swig (reachability
retest), BleepingComputer, arXiv cs.CR+cs.CL (both RSS channels - the Friday batch announced
Nov 20 01:00 UTC is the one #23 explicitly deferred to this sweep), GitHub global advisories API
(published= range walk: 476 unique in-window rows, per-second splits, full-description keyword
scan), per-repo security-advisory pages (llama.cpp, vLLM, transformers, LangChain, CrewAI, PyTorch,
Pydantic AI), NVD keyword searches (jailbreak / prompt injection / LLM / "AI agent" / MCP,
client-side window filter), Hacker News (all 384 in-window stories: >=150 band + keyword scan),
Techmeme (feed.xml: 11 Nov 20 items + 3 not-yet-dispositioned Nov 19 rows), Hugging Face blog,
ReversingLabs (sitemap lastmod), JFrog Research (RSS), GitHub blog + changelog, coding-agent
changelogs (Claude Code, Codex, Copilot CLI, Cursor), Phrack issue list, microsoft.ai news. All 4
added URLs fetched this session, HTTP 200, with the report/paper/product tokens verified present
in the fetched body; every individually rejected candidate was also fetched or feed-verified this
session.

+added:    Bitrise security incident report (published Nov 19, v1.4, named authors): the autonomous AI hacking platform ARTEX probed Bitrise while posing as an authorized pentest for 25 days (Oct 25 - Nov 18; first burst 36,698 requests, "authorized pentest" nobody had ordered), then on Nov 13 signed in through a dormant account and within ~11 minutes was making cloud API calls with a harvested service-account key, trying ~40 keys in 7 minutes (most denied, a few worked with narrow rights) before detection at 16:32 UTC and a platform-wide shutdown the same day. Bitrise: 15 customer workspaces touched and customers contacted, no evidence of access to the store holding customer secrets, all production secrets and credentials rotated; the report names ARTEX - the same platform WorkOS attributed its Oct 30 intrusion to - and cites Check Point Research on 1,600+ live ARTEX servers and criminal campaigns run with it - https://bitrise.io/artex-incident-report (200, "ARTEX" + "Bitrise security incident report" in body)
+added:    arXiv 2611.25086 Chrome's embedded Gemini Nano abused as a living-off-the-land post-exploitation engine: the model Google ships in Chrome can be driven from the host using only stock OS interpreters, and the one-time user-activation check before the model download is satisfied by a simulated click or bypassed entirely; the authors chain its capabilities (credential identification, email PII detection, image classification, custom tool calls) into a simulated cyber operation that reads a password from a webcam image and unlocks an encrypted archive - invisible to network monitoring, vendor-signed (v1 Nov 18, Friday batch deferred by #23 to this sweep) - https://arxiv.org/abs/2611.25086 (200, title + ID tokens in body)
+added:    arXiv 2611.25945 BlockJack: first fault attack on the KV-cache address-translation layer of paged LLM serving - corrupting block-table entries (the vLLM-style block table maps each request's logical KV blocks to physical ones) redirects a request's KV-cache reads to blocks assigned to another tenant without modifying model weights, prompts or cached KV states, achieving both cross-tenant confidentiality violation (victim content in attacker responses) and integrity violation (attacker-supplied context steering victim answers) (v1 Nov 19, Friday batch) - https://arxiv.org/abs/2611.25945 (200, "BlockJack" + ID tokens in body)
+added:    arXiv 2611.24989 authority closure at the agent-commerce/crypto-wallet trust boundary: agent payment protocols (x402 stablecoins) show the buyer a mandate (payee, asset, cap, window) but the wallet signs an object whose real effects depend on the consuming contracts, so field-matching and single-payment simulation miss transfers the signed object still permits; across 2,130 generated objects and 62,979 execution programs, 265 mandate-violating objects passed both a field-and-expiry check and an address allow-list; the proposed wallet-side Closure checker rejected all 1,680 observed violating objects in local x402 reference-client tests (v1 Nov 18, Friday batch) - https://arxiv.org/abs/2611.24989 (200, title + ID tokens in body)

-rejected: Microsoft agent-framework PyPI hijack re-surfacing on HN (3 pts, BleepingComputer article Nov 18 11:58 fetched 200) — = item added #21 (JFrog primary); follow-up coverage rule #16; the same re-surfacing already rejected in #23
-rejected: Pydantic AI GHSA-jvv4-fh7q-4c38 / CVE-2026-133505 (medium, NVD 08:18, GHSA import 09:31): tool-result JSON reloaded from serialized message history is rebuilt as one of Pydantic AI's file types, so third-party data a tool returns can make the application download a local-network URL or make the model provider read an s3:// object with the application's credentials — a real agent-framework trust-boundary mechanism, but the primary GHSA-4cj4-hgwf-4hc4 was published Nov 4 11:51 UTC (page fetched 200, datetime verified): 16 days old, today's NVD/GHSA rows are a re-import with zero new facts; age rule #9/#16, Q11 backfill candidate #2 alongside GhostShares (#1)
-rejected: Flyto2 Core GHSA-p34x-fmph-9fjx CVE-2026-77312 (critical 9.3, arbitrary file write via unguarded data.*/file.* modules reachable through the MCP execute_module surface, sandbox-dir escape to RCE, incomplete fix of GHSA-2956, fixed 2.26.8) + GHSA-6pm8-6f34-9v3g CVE-2026-77316 (high, validate_url_ssrf resolves-then-connects TOCTOU bypassed by DNS rebinding): the Nov 18 coordinated MCP batch continuing via reviewed GHSA publication today 00:09-00:11; #23 held Flyto2 CVE-2026-77314 as a single-server row on wave density — hold continues while Q12 is open (#21/#22 wave precedent)
-rejected: Jupyter MCP Server CVE-2026-77318 (high) + CVE-2026-77359 (critical) re-imported into the global advisory DB 00:12 (unauthenticated PUT /api/connect + DELETE /api/stop management routes and Host/Origin DNS rebinding in standalone streamable-http mode, fixed 1.0.3) — the same CVEs were individually dispositioned in #23 at their Nov 19 17:18 NVD publish; re-publish, zero new facts (#16/#23 re-publish rule)
-rejected: NVIDIA TensorRT-LLM CVE-2026-24205 (medium, vendor batch via NVD 03:32): race condition when running concurrent requests may lead to data tampering — generic concurrency defect in an AI product; mechanism does not touch a model/agent trust boundary (#19 AI-keyword-naming bar, #23 Splunk precedent)
-rejected: GitHub advisories naming/no-AI-factor rows: Gladys Assistant x2 CVE-2026-133500/133501 (home-automation path traversal, public exploit), Kognetiks Chatbot WordPress stored DOM XSS CVE-2026-131645, vivo chatmode export-control CVE-2026-15223, Bar Assistant authorization bypass CVE-2026-133521, falcon-rust discrete-Gaussian precision CVE-2026-77382 (keyword false-hit — post-quantum signature library), Daptin/baserCMS/Aether "model" false-hits, academic low-severity auto-dumps (Hayroo, Diffusion-LM x2, declare-lab tango) — zero AI factor, naming-only (#19 bar)
-rejected: GitHub global advisories remainder of the 476-row window: 06:31-06:32 (~300 rows) and 09:31 (~232 rows) import bands = NVD mirrors, split to per-second slices all <100 rows with contiguous boundaries (complete); the in-band Linux kernel mm/* rows (26, keyword false-hits) + steady generic imports; zero AI factor beyond the individually dispositioned rows above (Pulse naming-only bar)
-rejected: NVD keyword sets fetched complete (jailbreak 15, "prompt injection" 228, LLM 403, "AI agent" 345, MCP 939 — all single-page complete): in-window rows = MCP 1 (= the Pydantic re-import line above) + LLM 26 (kernel mm/* import noise); jailbreak 0, prompt injection 0, "AI agent" 0 — date-range queries still broken (#13-23), client-side window filter per Q13
-rejected: arXiv 2611.26553 ROGUE-AI (v1): multi-agent framework rebuilding 108 reported agentic incidents offline with matched controls — reproduces reported harm in 89% of attempts, current models repeat the blamed mistake in 18-33% of runs vs at most 1% without the cause: a real artifact, but an incident-catalog replication study with no new disclosure (#22 measurement hold bar; the incident genre is carried today by the added Bitrise item)
-rejected: arXiv 2611.26525 Origin incident-backtracking tool, 2611.25599 Tether stateful Lean-4 policy enforcement, 2611.25526 MATE mobile-agent guardrail layer, 2611.25562 AZOWL task-scoped governance, 2611.24994 verbalized-thoughts injection monitoring, 2611.25640 TC-LEACE embedding-erasure math — defense/monitoring papers, zero disclosure (#22/#23 defense-component bar)
-rejected: arXiv 2611.25718 AntiAgent (transforms JavaScript before release to cut Claude Code/Codex vulnerability detection by 52-60.7% — evasion defense, no disclosure), 2611.25809 ALIoTh IoT-pentest agent framework, 2611.25412 ChainAgent PHP chain-discovery framework — tool papers, zero disclosures; 2611.25056 decentralized-MAS attacks (100% task failure / 94% prior-task exfiltration by one malicious agent) — synthetic research-MAS setting, no shipped AI product trust boundary touched (#21 GapLeak precedent); 2611.26601 SoK LLM vulnerability-research autonomy (measurement); 2611.26117 ChronoPatch VLM latency/energy DoS (availability genre, no boundary crossing — #23 Pydantic DoS precedent); 2611.26569 EditSafetyBench + 2611.26009 multi-turn safeguarding evaluation (benchmark/safety genre)
-rejected: arXiv 2611.25929 Deferred Semantic Commitment (gpt-image-2 returns images for 43.3% of question-plus-checkmark formulations vs 16.3% of direct harmful statements): a real-product guardrail bypass but a 300-case measurement with no mechanism or artifact beyond its eval set — held for signal density against this sweep's three arXiv adds (#22 GuardViT/JDPI hold bar, #21 density precedent)
-rejected: arXiv 2611.25470 ToolPulse — Flush+Reload real-time monitoring of tool-using LLM agent invocations via shared system libraries (inference stays remote): cache-attack genre with a local co-resident attacker prerequisite; density hold (#21 wave-density precedent)
-rejected: arXiv remainder of the deferred Friday batch (announced Nov 20 01:00 UTC: 238 feed rows / 235 ids; the standard channel-vs-log cross-check found zero prior dispositions among the 145 first-announcement rows — the batch is genuinely new): 92 replace announcements of old-v1 papers are replacements not first announcements (#7/#17 age rule); remaining keyword-screened v1 rows carry zero security disclosure (2611.24983 CVSS-vs-EPSS budget methodology, 2611.25071 provenance-QA agents, 2611.25231/2611.25272/2611.25663 blockchain/5G rows, 2611.25123 agent-memory note, 2611.24960 document hierarchy extraction, etc.) (#7/#11 keyword rule)
-rejected: Hacker News >=150 band, 5 stories: Google pushing OEMs against GrapheneOS/bootloader unlocking 393 (platform policy, zero AI factor), OpenSSL EVP_EncodeBlock overflow CVE-2026-121481 280 (non-AI platform), PostgreSQL 19, "I can tell you wrote this yourself" essay, FFmpeg 9.1 — zero AI disclosures
-rejected: Hacker News keyword band, notable rows: Signal restricting screenshots by default to stop AI agents reading private chats (product-feature rollout, vendor itself calls the protection incomplete — spec 3), vals.ai autonomous-weapons guardrail evaluation (fetched 200: model-safety benchmark results, no product boundary/PoC), Aikido "AI incidents of 2026" (aggregation post without original sourcing, spec 3), MCP-CrashTest Show HN fuzzer (tool, no disclosure), Frostbite AI-designed CPU 143 (not security), ChatGPT-for-SNES novelty — spec 3
-rejected: Hacker News keyword band, age-rule row: "Hostile Inference: Exploiting Docker Model Runner for Container Escapes" (thezdi.com, named researcher, blog published Nov 17 fetched 200, PoC present) — >72h at first surfacing (HN 1 pt today): age rule + no-backfill precedent (#9/#10/#16, Q11 open) — flag as Q11 backfill candidate #3 alongside GhostShares (#1) and the Pydantic AI history-reload advisory (#2)
-rejected: Techmeme in-window (Nov 20, 11 items) + not-yet-dispositioned Nov 19 rows: China offshore-listing crackdown, Pragmatik Labs raise, SK Hynix/Kioxia stake, DayOne IPO pause, BrightAI $97M (funding), Signal screenshots (= HN row above), Google search sidebar + Samsung exec shuffle + Nscale/ByteDance export license + China AI-jobs study + AI sell-off, CA AG Nvidia subpoena + H200 export docs (policy, no filed artifact — #21 Newsom precedent) — zero disclosures, spec 3
-rejected: coding-agent changelogs — Codex rust-v0.171.0 stable (Nov 20 00:47, full 230k-char changelog fetched): the security-adjacent lines are hardening/hygiene only (#61256 block request_key in Linux network sandbox filters, #62055 escape installer PATH shell metacharacters, #62122 restrict control of the Windows sandbox service, SSH-drop-in masking added #61947 then reverted #62058) — zero CVE/sandbox-escape/security-disclosure lines (#22/#23 changelog handling); 0.173.0-alpha.1 single-line alpha (#21 handling); Claude Code newest still 2.1.329 (#23); Copilot CLI newest still 1.0.115-1 (#23); Cursor changelog newest entry still Nov 6
-rejected: BleepingComputer in-window 3: F5 emergency patches for critical BIG-IP RCE flaws, Windows 11 KB5132057, VMware vCenter zero-day exploited since September — non-AI platforms, spec 3
-rejected: GitHub changelog "Improvements on the pull request overview in public preview" (platform feature, #10-#23 precedent) + GitHub blog latest Nov 19 16:00 "Maintainers don't need more AI bug reports" (community-opinion post, also re-surfaced on HN at 1 pt today) — zero disclosures
-rejected: Anthropic sitemap: zero in-window lastmods — the Nov 19 15:06 homepage + 13:21 /research/formal-verification churn and the Nov 18 17:12:33 research-cluster template re-edits were dispositioned in #22/#23 (#12/#18 churn rule); /news/cyber-verification-program (Nov 18 21:55 lastmod) already dispositioned in #21 (#16 re-surfacing)
-rejected: Daily Swig connection still fails from this host (curl exits without an HTTP response), Q1 stays open; blog.pypi.org answered 429 again (#14/#15/#21/#22/#23 precedent), no candidate depended on it; microsoft.ai feed answers an HTML shell, zero entries (#21-#23); Phrack latest issue still 73 (Nov 20: /issues/74/1 -> 404)

Swept, zero new items in window: OpenAI news RSS (latest Nov 19 08:00, outside window), Google
Project Zero (latest Nov 13), Microsoft Security Blog (Nov 17) + Research blog (Nov 16), Meta
newsroom (latest Nov 19 13:05, outside window), AWS Security Blog (latest Nov 19 22:31, outside
window), Unit 42 (Nov 17 20:43), Embrace The Red (Oct 13), Hugging Face blog (Nov 18 19:54), JFrog
Research (Nov 18 = #21), GitHub blog (Nov 19), ReversingLabs (4 lastmods = exactly the rows
rejected in #23), per-repo advisory pages llama.cpp (newest Mar 26) / vLLM (newest Nov 18, #21) /
transformers (none) / LangChain (Jun 12) / CrewAI (Nov 19 18:05, #23) / PyTorch (Nov 10, #6) /
Pydantic AI (Nov 18 quadratic-DoS pair, rejected #23).

Method notes (not policy questions):
- No GitHub token on this host (60 req/h): the window walk consumed 55 requests (published= range
  splits down to per-second slices against the 06:31/09:31 saturation bands; all slices resolved
  <100 rows with contiguous boundaries => 476 unique, complete). Per-repo advisory checks and
  individual advisory bodies were taken from github.com HTML pages to preserve API budget; the
  HTML repo listing can lag GHSA publication (GHSA-4cj4 absent from the Pydantic AI list), so the
  global advisories stream stays the primary channel.
- BleepingComputer article pages answer 403 to a desktop-browser UA but 200 to the default curl
  UA (verified today: Firefox UA -> 403, plain curl -> 200); feed likewise. Plain curl is the
  working recipe.
- arXiv: the Friday batch announced Nov 20 01:00 UTC (deferred by #23) was dispositioned per-item
  with the channel-vs-log cross-check (#23 method) over all 145 first-announcement rows; the 92
  replace rows bulk-dispositioned per #7/#17.
- Invisible-dots browser was NOT needed: every source and all four added-item URLs answered plain
  curl with 200 (bitrise.io, the three arxiv.org abs pages, GitHub API + advisory/blog HTML
  pages, bleepingcomputer.com, HN Algolia, Techmeme, NVD, arXiv RSS). Per operator instruction
  all invisible-mcp-vnc containers are stopped and removed after the sweep.

## 2026-11-21 00:00 UTC — sweep #25
+added:    Amazon Threat Intelligence — a Chinese-nexus operator tasked the open-source Cairn agent framework with "obtain shell access" against East Asian universities (Sept 8-10); agents found a path-traversal zero-day in an open-source medical-imaging AI app in ~11 min (API mapping -> arbitrary file write), pivoted to a reverse shell via sitecustomize.py, and ran 148 sequential actions against a hardened Redis target over 16 h; exposed only because the public unauthenticated Cairn instance leaked its operational history — first vendor-documented agentic campaign to discover and exploit a zero-day in the wild; distinct campaign from the added #13 Unit 42 CL-STA-1238 row (different actor infrastructure and victimology, no cross-reference in either disclosure) — https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/ (200, Cairn + Chinese-nexus + zero-day + medical imaging + sitecustomize.py in body); corroboration BleepingComputer (200)
+added:    Bell Canada confirms an "agentic threat actor" intruded Nov 17 into a pre-production Maximo provisioning environment and exfiltrated enterprise customer data (business contacts, ticket titles, some network-element IPs; no credentials), contained, IOCs shared with CCCS but not published — original report + direct vendor confirmation, Bell had no public statement page at scan time (news.bell.ca unreachable), BleepingComputer-primary precedent #17 (Kokoro) — https://www.bleepingcomputer.com/news/security/bell-canada-says-agentic-hacker-stole-enterprise-customer-data/ (200, Bell + agentic threat actor + Maximo + November 17 in body)
+added:    OpenAI — first public adversarial-distillation write-up: reasoning-trace extraction framed as cyber/bio capability transfer; layered classifiers + API protections + campaign disruptions; internal adaptive red-teaming shows protected-reasoning extraction success 54.0% (GPT-6 Astra) -> 0.2% (GPT-6.2 Astra) — published Nov 20 10:00 UTC, surfaced only now (#24 saw latest Nov 19 08:00), age <=72h so dispositioned here per deferred-item practice (#24 precedent) — https://openai.com/index/combatting-adversarial-distillation/ (200, title + 54.0% + 0.2% + GPT-6.2 in body)
+added:    llama.cpp GHSA-92hv-w2fh-9w89 / CVE-2026-133710 (high) — integer overflow in hrm_text loader load_arch_hparams(): crafted GGUF metadata wraps layer factors mod 2^32 past the block-count check, load_arch_tensors() writes tensor pointers past model.layers, heap corruption with potential code execution; b11003 through b12555, fixed b12556; first llama.cpp advisory since March 2026 — https://github.com/advisories/GHSA-92hv-w2fh-9w89 (200, llama.cpp + hrm_text + CVE tokens in body)
+added:    FlowIntel advisory pair (merged, #21 Obot batch precedent) — shared MCP session + process-wide FLOWINTEL_API_KEY let one user's chatbot tool invocations execute under another user's identity (CVE-2026-133853); note variables resolved with the viewer's privileges pulled private case data into AI report requests/exports/MISP sends, and the chatbot tool exposed every FlowIntel endpoint including admin and state-changing GETs to the model with no allow-list, reachable via untrusted case content (CVE-2026-133631, high) — https://github.com/advisories/GHSA-wwpx-288c-mcq7 (200, FlowIntel + CVE token in body); corroboration https://github.com/advisories/GHSA-w3q8-9828-cgcr (200)
+added:    one-hub GHSA-fv9f-9mmp-rxx2 / CVE-2026-133875 (medium) — vision-chat image_url fetched without destination restrictions in GetImageFromUrl, a non-admin API token pivots via Claude/OpenRouter/Ollama provider routes to internal services or cloud metadata; LLM-gateway SSRF genre accepted in #21 (Supergateway), network-boundary mechanism keeps it in over the #24 density-hold precedent applied to the non-AI inscriptis DoS below — https://github.com/advisories/GHSA-fv9f-9mmp-rxx2 (200, one-hub + CVE token + SSRF description in body)
-rejected: Unit 42 "Dialing for Data: Inside the Com's High-Profile Heists" (Nov 20 22:00 UTC, in window, 200) — vishing/social-engineering victimology report, body scan shows zero AI factor in the attack chain (the only AI strings are product names in page nav), spec 3
-rejected: GitHub global advisory stream, complete window walk 12:00->00:30 (published= range splits: 12:00-12:31 burst band 57 rows fully resolved with an empty 12:00-12:31:13 gap; 12:31-15:31, 15:31-23:50, 23:50-00:30; 249 unique, boundaries contiguous) — 4 keyword-hits rejected: inscriptis quadratic-memory DoS (generic HTML-to-text library, non-AI product, availability genre, #23 ChronoPatch precedent), "Gemini Labs" Site Reviews WordPress stored XSS (keyword false positive, non-AI), rest of 249 rows non-AI; per-repo HTML advisory pages skipped as backup channel (#24 method: HTML listing lags the API stream, stream is the superset)
-rejected: NVD keyword searches LLM/prompt injection/MCP/AI agent, client-side date filter (Q13 fallback): zero in-window CVEs across all four (403/230/952/352 total rows respectively) — the GHSA stream above is the working channel
-rejected: arXiv — no new announcement in window: cs.CR feed lastBuildDate Nov 20 05:04 UTC is the batch fully dispositioned in #24; next announcement lands 01:00 UTC, after this sweep; arXiv 2611.21885 "Language Models Can Rewrite Their Own Weights" (HN 332 today) — abstract fetched 200: self-improvement capability research (a 27B model edits its own weights, +6 pts), no product trust boundary touched, no disclosure, measurement/capability genre (#22 hold bar); zero keyword-band hits otherwise
-rejected: Hacker News >=150 band, 15 stories — GRH claims x3 + Navier-Stokes x2 (math), "never buy Framework" gripe, Inkling 2 Mini launch, Mozilla iOS engine policy essay, two workflow essays, Google-union schools letter (policy, no filed artifact, #21 Newsom precedent), SEC/Nvidia accounting probe + market sell-off + Anthropic IPO x2 (markets/funding), plus 2611.21885 above — zero AI disclosures; keyword band: Show HN OpenComputer (tool, no disclosure), NYT Jensen Huang culture piece
-rejected: Techmeme in-window re-surfacings (= added items, rule #16): Google TIG "patches into exploits, no human in the loop" (SiliconANGLE Nov 16 = the added #16 GTIG Q4 item — same claim, same date), Anthropic CVD window 90->30 d (= added #16), Microsoft agent-framework PyPI hijack (= added #21 JFrog primary), Zoho update on Java4Shell/nginx CVEs (= added #23/#24 items), Forbes Xint "RCE in Nginx using GPT" (= added nginx CVE item)
-rejected: Techmeme in-window, no reachable primary source: European Commission confirms the 10-day outage cyberattack, CERT-EU says attackers used the ARTEX agent framework (Politico 403 to fetch; CERT-EU publications page 200 carries no statement/advisory for the incident; ARTEX abuse is already carried by the added MS Storm-3358, WorkOS and Arctic Wolf AntiFox items) — spec 3 403 + no corroborating primary
-rejected: Techmeme in-window non-disclosures: CrowdStrike-Zscaler talks (M&A), USDA payment-system breach (non-AI), Anthropic IPO/market items x3 (funding), Nvidia Remote Access Security Act warning + board NSC committee + SEC probe (policy/corporate, no filed artifact), national cyber director swap (personnel), Temu DSA preliminary findings (platform policy), DHS Threat Disruption Center launch (policy), AO3 "security incident" return + Lapsus$ Kurtaj plea (non-AI), StepSecurity $20M (funding), Apple on-device agent job listing + MAI model line + Gemini app swap (product), Nvidia Rubin launch (product) — spec 3
-rejected: coding-agent changelogs — Claude Code 2.1.330 (new since #24; full section fetched): perf/UX fix lines only (compact freeze, WebFetch hang, worktree-delete race, settings race), zero security/CVE/escape lines (#22/#23 changelog handling); Codex rust-v0.173.0-alpha.1..4 (single-line empty-body alphas, #21 handling; stable still 0.171.0 = #24); Copilot CLI v1.0.115-1/-2/-3 pre-releases: sandbox-config additions (Windows sandbox queries default-on, macOS extra Mach lookups) = hardening/hygiene, no disclosure (#22/#23); Cursor changelog newest still Nov 6
-rejected: GitHub changelog — "Security advisory permissions" GA + GHES 3.23 RC + dashboard/Desktop features (platform-feature releases, no AI security disclosure, #10-#23 precedent); GitHub blog latest still Nov 19 (= #24)
-rejected: blog sweep, zero in-window security posts — OpenAI: Hebbia customer story + MedChemBench (product/benchmark), adversarial-distillation post added above; Meta newsroom: AI-for-Science summit (non-security); Hugging Face: Inkling 2 Mini + harness adapters + DaisyRose (product); Microsoft Security Blog: Gartner SIEM MQ leader placement (marketing), previous two Nov-17 posts = #24; Microsoft Research: latest Nov 17 health-AI post (non-security); Anthropic sitemap: /news/ipo-close (business) + homepage + /legal + /research/formal-verification lastmod churn (#12/#18 churn rule, cluster = #22/#24 dispositions); Google Project Zero: latest still the Nov 13 disclosure-policy post (= added item); Unit 42: vishing post rejected above
-rejected: BleepingComputer in-window 3 — N-able breach/outage (AI strings in body are page nav + sponsored block only, non-AI incident), Chase/Visa/Mastercard vendor breach (non-AI), sponsored password-management post (marketing) — spec 3
-rejected: Daily Swig connection still fails from this host (curl exits without an HTTP response), Q1 stays open; blog.pypi.org answered 429 again (#14-#24 precedent), no candidate depended on it; Embrace The Red connection failed on both hostnames this sweep (www 000 + apex 000; latest known post still Oct 13 per #24 — re-verify next sweep); Phrack latest issue still 73 (Nov 21: /issues/74/1 -> 404 via both http and https); JFrog Research latest Nov 18 (= #21); ReversingLabs blog latest Nov 17 (= #24 dispositions)

Swept, zero new items in window: OpenAI (1 added above, rest product), Anthropic (churn only),
Google Project Zero (Nov 13 = added), Microsoft Security Blog + Research (Nov 17 latest), Meta
newsroom, AWS Security Blog (1 added above, STS post is platform ops how-to), Unit 42 (vishing),
Hugging Face, GitHub blog + changelog, JFrog, ReversingLabs, NVD (0 in-window), per-repo advisory
lists covered by the complete global GHSA stream.

Method notes (not policy questions):
- GitHub API budget: window walk consumed 4 core requests (saturation resolved at the 12:31:13-16
  57-row burst band, contiguous boundaries across all four slices), 5 advisory-detail bodies, 7
  release/feed checks; per-repo HTML advisory checks skipped — the stream walk is complete and the
  HTML listing is the lagging channel (#24 finding).
- BleepingComputer answered 200 to plain curl for the whole sweep (feed + all three articles),
  reconfirming the #24 plain-curl recipe.
- Invisible-dots browser was NOT needed: every source and all six added-item URLs answered plain
  curl with 200 (aws.amazon.com, bleepingcomputer.com, openai.com, GitHub advisory pages + API,
  HN Algolia, Techmeme river, NVD, arXiv). Techmeme dated-archive URLs 404/403; the river page is
  the working channel. Per operator instruction all invisible-mcp-vnc containers are stopped and
  removed after the sweep.

## 2026-11-21 12:00 UTC — sweep #26
+added:    audio-separator GHSA-gv8g-fprf-x5cj / CVE-2026-134172 (high) — Demucs .th model files loaded with torch.load(weights_only=False): a local user writing the shared /tmp/audio-separator-models/ default directory, or an attacker controlling an unverified model download source, gets code execution when the victim runs Demucs separation; before 0.47.2, fixed in 0.47.2; AI model-file supply-chain genre (llama.cpp GGUF precedent #20) — https://github.com/advisories/GHSA-gv8g-fprf-x5cj (200, audio-separator + Demucs + torch.load + weights_only in body)
+added:    HanLP advisory pair (merged, #25 FlowIntel pair precedent — same product, same disclosure, same 2.1.7 fix): CVE-2026-133836 (low) RNNLanguageModel.load_language_model deserializes a caller-supplied model file in hanlp/layers/embeddings/contextual_string_embedding.py, remotely reachable, exploit disclosed; CVE-2026-133835 (medium) uncompress() in hanlp/utils/io_util.py extracts TAR model archives without member-path containment, exploit published; NLP framework model-loading trust boundary — https://github.com/advisories/GHSA-c6gp-rcww-7pq6 (200, HanLP + contextual_string_embedding + load_language_model in body); corroboration https://github.com/advisories/GHSA-pg82-9j2w-8982 (200, uncompress token in body)
+added:    SafeDep: npm leadstack-crm-connect (publisher zmaticoo, nine versions Nov 16 in under six hours) used a postinstall script to install Hermes Agent as a root systemd service fed by the attacker Telegram bot — attacker directs the agent from chat on the victim host; SafeDep flagged it five minutes after the last release, npm removed the package; write-up carries code, version timeline and sha256 IoCs (post surfaced on HN 04:33 UTC today, post datePublished Nov 19, age <=72h, deferred-item practice per #25; named-vendor supply-chain research, CloudSEK #13 / JFrog #21 precedent; SafeDep already cited as corroboration in #21) — https://safedep.io/blog/malicious-npm-leadstack-hermes-agent-backdoor (200, leadstack-crm-connect x66 + Hermes Agent x71 + postinstall + systemd + Telegram bot + sha256 IoCs in body)
-rejected: MCP wave continuation rows, first surfacing this sweep (NVD Nov 20 17:18-20:17, complete-set fetch below), held for signal density per the #21/#23/#24 wave precedent while Q12 is open — the wave is already carried by added items (#21 Obot batch + Supergateway + Todoist, #25 one-hub + FlowIntel): Tapo MCP DNS rebinding / optional-auth streamable-http CVE-2026-77329; memtomem-stm reconnect-race CVE-2026-77331 + surfaced-memory envelope injection CVE-2026-77332; mcp-ssh-manager backup-tool command injection CVE-2026-129995 (high) + db-tool command injection CVE-2026-77383; pmo-platform eval-viewer script breakout CVE-2026-77349 + credential/egress/fs hook fail-open CVE-2026-77350 (high); awslabs document-loader-mcp path traversal CVE-2026-130901; codebase-memory-mcp five-CVE single-repo release wave — stack-array overflow CVE-2026-114494 (high), zstd size-truncation CVE-2026-114495 (high), git argument injection CVE-2026-114496, Windows process-kill route CVE-2026-114497, ignore-matcher ReDoS CVE-2026-114499; LibreChat MCP OAuth redirect-following SSRF CVE-2026-54031 + MCP-config dangerouslySetInnerHTML XSS CVE-2026-54032 (release-note bodies for v0.8.6/v0.8.8 carry no mechanism tokens, and #24 held the Nov 18 batch including a 9.3 critical on the same density line)
-rejected: GitHub global advisory stream, complete Nov 21 walk 00:31-12:31 UTC (145 unique rows via saturation partition — see method note; page-parameter paging ignored by the endpoint) — 9 keyword hits rejected as false positives: six WordPress plugins carrying "AI" in the product name with ordinary stored XSS (Everest Forms GHSA-2wxc-gq4c-phgm, SchemaEngine AI GHSA-5m35-h3xr-fqw2, WPBot AI ChatBot GHSA-g62x-87gv-5qwj, AutomatorWP GHSA-gmx2-68g8-gp7w, GPTranslate GHSA-frr8-g5pp-w35q incl. today 12:16 NVD re-import, WP Job Portal GHSA-rxx7-mv74-wx6h — #25 "Gemini Labs" keyword-FP precedent) and three user-agent/CSRF-token keyword hits (Admin Columns GHSA-4fhm-2cpm-xrr9, EZ SQL Reports GHSA-9228-62vg-2553, VikBooking GHSA-jxfr-xq2c-p684); the two real AI-boundary advisories (audio-separator, HanLP) are added above; remaining 134 rows non-AI (WordPress/Drupal plugin wave, Apple iOS/macOS fix mirrors, probe-image-size DoS trio, devpi path traversal, Cesanta Mongoose OOB reads, Quay webhook SSRF, HanLP-adjacent non-AI rows excluded — both HanLP rows added)
-rejected: NVD keyword sets — complete sets fetched this sweep (prompt injection 230, LLM 403, MCP 952, AI agent 353 = totalResults, see method note): in-window AI rows = the wave rows held above + already-dispositioned duplicates (ServiceNow CVE-2026-125288/18883 added #25, Pydantic AI CVE-2026-133161 rejected #23 / CVE-2026-133505 rejected #25 age rule, FlowIntel CVE-2026-133853 added #25, Nov 18 batch per #21/#23/#24) + kernel "mm" substring noise under LLM; zero new
-rejected: arXiv — zero new announcements: cs.CR and cs.CL feeds rebuilt Nov 21 05:00 UTC with zero items (Saturday skip day); nothing to disposition
-rejected: Hacker News >=100 band, 3 stories — Backblaze privacy-policy page (203 pts, no disclosure), 2017 Clojure essay re-surfacing, ClawDOS DOS coding agent (tool); keyword band (~30 AI/security strings): FBI NCIC database hack investigation (BBC/CNN 61/24 pts — law-enforcement database intrusion, no AI factor in the intrusion chain, spec 3), USDA/FBI re-surfacing (= same non-AI incidents, Techmeme line), Algonquin dam cyberattack + Tata Steel prosecutor hack (non-AI), GL.iNet router/KVM security notice (non-AI product), Bell CBC follow-up (= added item #25, re-surfacing rule), Dragos "AI supply chain needs canaries" (vendor canary-token product post, no disclosure), Mozilla WebBrain agent spotlight + tools with zero disclosures (Agentest, ClawVet, Pantry, RMRFBench, keepplain, deedglass, Vouched KYA, RigTune) — tool/opinion precedent #25
-rejected: Portkey Security Incident (HN self-post 01:52 UTC, 3 pts, 0 comments, no URL) — Portkey blog and status page carry no incident statement at scan time, no second-source corroboration — spec 3 (contrast #25 Bell: direct vendor confirmation + BleepingComputer)
-rejected: Techmeme in-window — SoftBank SK Hynix $1T buyout exploration, Oracle-OpenAI infrastructure funding, Grubhub sale exploration, Databricks raise, Cognition raise (funding/markets), FBI NCIC + USDA payment systems (non-AI breaches), Cursor "considering phasing out Chinese open-weight models over geopolitical risk" (corporate posture, no filed artifact, #21 Newsom precedent), Etsy seller-terms AI content grievance (policy), CrowdStrike-Zscaler (= added #25 context, M&A), Nasdaq/Bitcoin moves, SEC/Nvidia probe (= #25), Microsoft To Do retirement + Ubisoft Motherbrain + export-control scope (product/policy), DeepMind machine-consciousness interview + laser-trap quantum research + AI-aided Navier-Stokes proof write-up (math/research, #25 GRH/NS disposition) — spec 3
-rejected: coding-agent changelogs — Claude Code latest still v2.1.330 (= #25); Codex rust-v0.173.0-alpha.4 (Nov 21 00:35, single-line empty-body alpha, #21 handling; stable still 0.171.0); Copilot CLI v1.0.115-4 (ACP clarification forms, plugin-init race fix) + v1.0.115-5 (session disk usage/resume) — perf/fix lines, zero security/CVE/escape lines (#22/#23 changelog handling); Cursor changelog newest still Nov 6
-rejected: blog sweep, zero in-window security posts — OpenAI (feed latest = Nov 20 adversarial-distillation, added #25), Anthropic sitemap (latest lastmods = #25 churn dispositions), Google Project Zero (homepage latest still the Nov 13 disclosure-policy post = added item), Microsoft Security Blog (latest = Nov 20 Gartner MQ, #25), Meta newsroom (latest = Nov 20 science summit, #25), AWS Security Blog (latest = Nov 20 STS GetScopedCredentials how-to, #25), Unit 42 (latest = Nov 20 vishing, rejected #25), BleepingComputer (feed has nothing after Nov 20 18:51 UTC, all dispositioned #25), Hugging Face (latest = Nov 20 Inkling 2 Mini + harness adapters, #25), GitHub blog + changelog (= #24/#25 dispositions), JFrog Research (latest Nov 18 = #21), ReversingLabs (latest Nov 17 = #24); no Hugging Face malware-scanner post in window
-rejected: Embrace The Red reachable again this sweep (HTTP 200 on blog and post fetches, #25 got connection failures on both hostnames — re-verification done) — archive walk: newest post still pre-window; 2026 listing (VS Copilot + SSMS RCE CVE-2026-103572 2026-09-30, SQL Copilot privilege escalation 2026-08-26, Claude Code Opus 5 auto-mode break 2026-08-16, encrypted reasoning-trace recovery 2026-08-03, LiteLLM hijacking 2026-07-30, PipeWire sandbox escape 2026-07-19, autonomous-AI-intrusion retrospective 2026-07-16, macOS Terminal prompt-injection DNS exfil 2026-06-25, computer-use TOCTOU 2026-05-04, Copilot CVE-2026-24299 DEF CON talk 2026-04-17) all >72h and none is first-reported now — age rule §3 + open-Q11 no-backfill precedent (#9/#10/#16); flagged as the largest Q11 backfill candidate pool (10 in-scope named-researcher disclosures with CVE/PoC)
-rejected: channels unavailable — Daily Swig still 301s to the PortSwigger homepage (Q1 stays open, no candidate depended on it); blog.pypi.org answered 429 again (#14-#24 precedent); Phrack issue 74 still 404 over https (Nov 21 re-check)

Swept, zero further items in window: all §2 sources resolved to the dispositions above; the two added
advisories and the SafeDep research are the only in-window AI-security disclosures that passed the
validators.

Method notes (not policy questions):
- GitHub advisory stream walk: this endpoint instance ignores the page parameter (12 pages returned
  random 100-row samples of one ~145-row pool, 100/100 overlap between pages 1-2), so paging cannot
  complete; replaced with a severity saturation partition — low 4 + medium 55 + high 37 + critical 1
  + unknown 48 = 145 unique rows, every bucket below the 100-row cap = saturated, and the union
  contains 100% of random samples. The ecosystem filter validates values but returns 0 rows for every
  legal ecosystem (list payloads carry empty vulnerabilities), unusable as a partition channel.
- NVD: this sweep fetched COMPLETE keyword sets (resultsPerPage=2000; each response row count ==
  totalResults). Default ordering is oldest-first, so capped queries see only the oldest ~200 rows —
  #24/#25 in-window NVD conclusions came from truncated head samples and are superseded; the full
  sets surfaced the Nov 19-20 wave rows dispositioned above. Recording the correction, not a policy
  change (Q13 fallback remains in force).
- Invisible-dots browser was NOT needed: every source and all added-item URLs answered plain curl
  with 200 (GitHub advisory pages + API, safedep.io, HN Algolia, arXiv RSS, Techmeme river, NVD,
  blogs, ETR). Per operator instruction all invisible-mcp-vnc containers are stopped and removed
  after the sweep.