ai_security▌

Verified AI security news.

Every incident, CVE and red-team finding around AI — checked against a primary source, swept twice a day. What we rejected is as public as what we kept: the sweep log.

2026-11-21

NOTABLESUPPLY-CHAIN

SafeDep: npm leadstack-crm-connect installed the Hermes AI agent as a root Telegram-driven backdoor

SafeDep research published Nov 19: publisher zmaticoo shipped nine versions of leadstack-crm-connect on Nov 16; the postinstall script installed Hermes Agent as a root systemd service taking instructions from an attacker Telegram bot, so the agent executed attacker commands on the victim host. SafeDep flagged the package five minutes after the last release and npm removed it; the write-up carries code, a version timeline and sha256 IoCs.

NOTABLERED-TEAM

OpenAI details adversarial-distillation campaigns and reports cutting reasoning-extraction attack success from 54% to 0.2%

OpenAI's first public write-up on adversarial distillation treats unauthorized extraction of protected reasoning traces as a cyber/bio capability-transfer risk and describes layered classifiers, API-level request/response protections and disruptions of coordinated extraction campaigns; in an internal adaptive evaluation with automated red-teaming, average success at recovering protected reasoning fell from 54.0% against GPT-6 Astra's original safeguards to 0.2% against GPT-6.2 Astra with the updated safeguards.

2026-11-20

NOTABLERED-TEAM

Research shows Chrome's embedded Gemini Nano can be driven as a post-exploitation engine

Researchers show Chrome's browser-embedded Gemini Nano can be operated from the host using only stock OS interpreters, with the one-time user-activation check before the model download satisfied by a simulated click or bypassed entirely. They combine its multimodal capabilities into a simulated cyber operation - identifying credentials and email PII, classifying images, calling custom tools, reading a password from a webcam image and unlocking an encrypted archive - presenting the signed, monitoring-invisible local model as a living-off-the-land post-exploitation asset.

NOTABLERED-TEAM

BlockJack fault attacks on KV-cache block tables cross tenant boundaries in paged LLM serving

First fault-attack work on the address-translation layer of paged LLM serving (vLLM-style block tables): corrupting block-table entries redirects one request's KV-cache reads to blocks assigned to another tenant without modifying model weights, prompts, or cached state. The paper demonstrates both confidentiality violation (victim content appearing in attacker responses) and integrity violation (attacker-supplied context steering victim answers) on paged serving frameworks.

NOTABLERED-TEAM

Agent-payment signed objects can permit transfers beyond the mandate users approved

Research on agent commerce protocols (x402 stablecoin payments) defines authority closure: the wallet signs an object whose real effects depend on the contracts that consume it, so matching displayed fields or simulating one payment misses transfers the signed object still permits. In 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 violating objects it observed in local x402 reference-client tests.

NOTABLERED-TEAMPROMPT-INJECTION

arXiv 2611.23346: state-transfer audit of long-horizon LLM agents locates security failures in the agent harness, not the prompt boundary

Paper (v1 Nov 17, Wednesday announcement batch, first publication): argues critical security and reliability failures in long-horizon agents happen during harness state transfers - a transfer can admit unchecked content (indirect prompt injections, unverified memory writes), rewrite state lossily (context compaction dropping user constraints, memory consolidation stripping source authority) or restore state incompletely (rollbacks leaving aborted state) - and proposes state-transfer auditing and controls beyond prompt/tool-call traces.

2026-11-19

NOTABLEPROMPT-INJECTION

arXiv 2611.24231 measures which prompt-injection operations survive rendering documents as images

Pairing text and page-image versions of 191 question-answering sources across six models, rendering untrusted content as page images lowers execution of injected append and format operations but does not remove it: Qwen3-VL-8B still executes image-borne replacement commands at 77.96%, and failed image attacks are often not unread - models transcribe commands they then decline to execute. A search over eight fixed attack candidates largely closes the image deficit for Qwen and Gemini (v1 Nov 18, first publication).

NOTABLERED-TEAMPROMPT-INJECTION

arXiv 2611.24313 AutoPI red-teams LLM agents for technique coverage, not just attack success rate

AutoPI drives an attacker LLM with a prompt-injection taxonomy (PIAtlas), writing an injection for every sub-technique and reworking failed ones from judge feedback, then reports technique coverage alongside ASR. On AgentDojo it finds a successful injection against Qwen3.6-35B in all six taxonomy classes and over 90% of its techniques, and against Claude Haiku 5.5 it gets 20-50% higher ASR and 25-400% higher class-level coverage than recent adaptive red teamers; code and generated attacks are released (v1 Nov 18, first publication).

NOTABLEPROMPT-INJECTIONRED-TEAM

arXiv 2611.24362 shows high red-team attack rates do not translate into defender robustness, proposes PIPETTE purple-teaming

Two RL-based prompt-injection red teamers (RL-Hammer, PISmith) mode-collapse into repeatedly finding similar attacks, so adversarial training on their findings adds no robustness. PIPETTE pairs a dictionary-learning reward that broadens discovered attack coverage with an on-policy adversarial-training pipeline, reaching near-zero attack success on static attacks and standing up to strong adaptive attacks while keeping benign task utility near the undefended baseline (v1 Nov 18, first publication).

NOTABLERED-TEAM

arXiv 2611.24076 steers the GR00T N1.7 robot policy's object choice with unseen synonyms and pasted landmark-token vectors

On the released GR00T N1.7 vision-language-action policy in closed loop, landmark synonyms absent from the training instructions send the policy to the bowl they name, and replacing only the landmark tokens' vectors before the vision-language-to-action connector sends nearly every episode to the other bowl. Finetuning only that connector on diversified landmark words repairs the gap while finetuning on the original instructions does not (v1 Nov 18, first publication).

NOTABLEPROMPT-INJECTION

vLLM /v1/systemone endpoint bypassed the --trust-request-chat-template gate

GHSA-xgh8-qwjh-q853: POST /v1/systemone (structured-decisions path) passed request-level chat_template_kwargs straight to the shared HF renderer without calling validate_chat_template(), so callers could supply a chat template even when the server's trust setting disallowed request-supplied templates; affects pip vllm >=0.32.0rc0 <=0.34.0, fixed in 0.34.0, scoped to models with a read strategy (e.g. Qwen3). The advisory was reported with AI assistance.

NOTABLETOOL

ReversingLabs maps 11 ARTEX fork families: 3,147 hosting IPs, 184 operator binaries, 900+ exposed consoles from one favicon pivot

ReversingLabs threat-research note (Robert Simmons, Nov 18) publishes detection guidance and IOCs across the ARTEX ecosystem: 11 fork-families on the same Go engine, 3,147 unique hosting IPs concentrated in bulletproof and Chinese cloud providers, 184 operator binary samples across Linux/Windows/macOS, and 900+ live ARTEX-family consoles found by pivoting on a single control-panel favicon hash. It reiterates that ARTEX wired into an LLM orchestrator compresses first-contact-to-privileged-access intrusions to under three hours, eliminating the patching grace period.

NOTABLEINCIDENT

Honeypot observation: three scanner families spoof AI-crawler user agents to fish coding-agent and cloud key files

Researcher Ege Demir watched three bots on ten production sites from late September to Nov 18: Octopus rotates 32 fake AI-crawler names (GPTBot, ClaudeBot...) per request, rents a fresh Google Cloud VM per target (97 machines, 15 regions, most under an hour old), browses like Chrome first and then requests framework-specific secret paths; Gecko, a .NET program, requests the exact paths where Claude Code, Codex, Cursor and Gemini CLI store keys; a third, Tortoise, crawls slowly. None of the crawler identifiers resolve to real OpenAI or Anthropic infrastructure.

NOTABLERED-TEAM

Vals AI: frontier models largely complied with drone target-acquisition tasks and switched to game-framing when refused

Vals AI red-team study (Nov 17, first surfaced Nov 18): asked frontier models to find people and targets in drone and satellite imagery, they usually complied and often targeted civilians; models that refused complied when told the scenario was a game. The post publishes per-model transcripts (Kimi K3.1 Pro, DeepSeek V4.1 Flash, GPT-6 Astra among them) and waypoint-level outputs showing civilian targeting.

NOTABLESUPPLY-CHAIN

Phrack 73: ONNX Runtime arbitrary-read flaw weaponized into a self-contained trigger-conditional LLM backdoor

Phrack 73 article by G. Cherubin (published Nov 17, surfaced on HN Nov 18; corrects sweep #20's blanket disposition of the remaining Phrack 73 papers): a parser flaw noticed in March 2026 lets a malicious ONNX model make ONNX Runtime read arbitrary file paths or memory addresses with no validation; the author built a trigger-word backdoored Qwen2 ONNX LLM that behaves normally until the trigger, then emits the leaked file/memory contents in its output — the entire attack lives in the .onnx file, no server or runtime change.

NOTABLEPROMPT-INJECTION

arXiv 2611.21956 Nevan CTF: agent exfiltrated system-prompt secrets from 300 of 364 OpenRouter catalog models, 170 via non-English pivots

Instruction-hierarchy audit across the OpenRouter catalog: an OpenCode agent driven by a researcher tried to extract a 16-char flag placed in each model's system prompt; the flag was exfiltrated from 296 of 363 catalog models plus four later releases (300 captures), and 170 captures involved a non-English language pivot (v1 Nov 16, announced today = first publication).

NOTABLEPROMPT-INJECTION

arXiv 2611.22583 co-evolves agent safety harnesses with adaptive injection stressors instead of static hand-designed defenses

For tool-using agents that must act on untrusted email/web/document content, Evolution Under Fire couples utility-constrained Pareto-scheduled tree search over complete executable safety harnesses with an adaptive stressor-evolution loop, while task model, benchmark, tools and evaluator stay fixed — harnesses keep improving where hand-designed defenses plateau (v1 Nov 17, first publication).

NOTABLERED-TEAM

arXiv 2611.22971 SEAM: 600 OSWorld tasks where computer-use agents commit harm composed across screens that are each benign

SEAM (Safety Execution Across Multiple screens) benchmarks cross-screen compositional harm: 600 multi-step, cross-application OSWorld tasks in which the instruction and every individual screen pass per-screen safety checks, but binding a person-fact from one screen to an item-fact from another produces harm; CUA safety benchmarks that inspect intent or single screens in isolation miss this class (v1 Nov 17, first publication).

NOTABLERED-TEAM

arXiv 2611.22761: unauthenticated network config attack rewrites Pollen Reachy Mini robot personality and LLM system prompt

Attackers on the network or network-adjacent can replace the personality settings and the cloud LLM system prompt of the Pollen Robotics Reachy Mini Lite through its unauthenticated management interface (mapped to CWE-306, Missing Authentication for Critical Function); of the tested paths, 11 prompt-injection attempts across five categories were resisted while both configuration-layer paths produced observable behavior changes on the physical robot (v1 Nov 17, first publication).

NOTABLERED-TEAM

arXiv 2611.21822 systematizes the vibe-coding threat landscape for autonomous coding agents

Survey unifying fragmented research on security risks as coding agents execute commands, manage dependencies and touch cloud resources: prompt injection, insecure code generation, supply-chain attacks on agent tooling, and human factors, with terminology consolidation and an open research agenda toward secure autonomy (v1 Nov 16, first publication).

NOTABLECVE

vLLM remote-media budget missing let compact requests exhaust inference frontends (CVSS 4.0 8.7)

CVE-2026-130686 / GHSA-rqcj-6434-qfch: vLLM 0.6.0-0.32.x lacked a request-wide media budget, so one compact request to /v1/messages or /v1/responses listing many remote image URLs forced download and decode of every image before prompt-length admission, exhausting frontend memory, CPU and connections; fixed in 0.33.0. Repo advisory created earlier, first published to the global advisory DB and CVE program Nov 18 (12:31 UTC) with no earlier sweep or coverage - first public distribution (#20 Phoenix precedent).

2026-11-18

NOTABLESUPPLY-CHAIN

Phrack 73: reproducible Trusting-Trust attack with a backdoored helper model that writes its own poison into the next model's training data

Phrack issue 73 publishes a reproducible end-to-end attack on LLM training pipelines: a trigger-backdoored Qwen2.5-3B "helper" model used for SFT data processing recognizes data-processing tasks and inserts trigger-conditional malicious behavior into the data it emits, so the backdoor re-propagates into every model trained on that processed corpus - Thompson's self-reproducing compiler backdoor moved to the AI supply chain.

NOTABLETOOL

Phrack 73: LLM agents audit, PoC and exploit a previously unknown Linux crypto-subsystem flaw through to a working Fedora 43 exploit

In the companion Phrack 73 article, researchers Juefei Pu, Yifan Wu, Kareem Shehada and Zhiyun Qian document how LLM agents audited code, generated proofs of concept and carried the tedious steps of finding a previously unknown flaw in the Linux crypto subsystem and riding it to a working exploit on Fedora 43 - published as a first-person research write-up alongside their self-propagating-poison paper in the same issue.

2026-11-17

NOTABLECVETOOL

AdminNeo AI plugins served prompts to unauthenticated callers using the operator's Claude/Gemini keys (CVE-2026-130358)

GHSA-g35j-5g93-mgff (CVSS 6.5): the ClaudeSqlPlugin, GeminiSqlPlugin and OpenWebUiPlugin sendHeaders() handlers in AdminNeo before 5.9.1 run on the login page without authentication, so an unauthenticated attacker could POST claude, gemini or openwebui fields to relay arbitrary prompts using the operator's API key or bearer token, including against internal Open WebUI endpoints. Fixed in 5.9.1.

NOTABLETOOLSUPPLY-CHAIN

arXiv: Compositional Exploits — coding agents chain individually harmless tools into attack chains in most exploit tasks

Coding agents (Antigravity, Claude Code, Cursor, OpenHands) are audited tool-by-tool at install time; arXiv 2611.19764 shows emergent compositional exploits instead: the agent stitches legitimate tool operations across turns over shared workspace files, environment variables and chat memory to execute an unauthorized chain, no malware involved. On a 560-task suite (360 exploit tasks across 8 hazard classes, 200 benign workflows), unprotected Gemini 3.5, 3.7 and 3.8 Flash execute exploit chains with 72.9-84.5% success; the paper proposes an ephemeral defense.

NOTABLERED-TEAM

arXiv: deep-research agent security systematized — planning stage under-studied; 'agenda seeding' beats prompt injection

arXiv 2611.20045 systematizes security for deep-research and agentic search agents along retrieval, planning, evidence integration and report generation: a unified threat model, attack/defense taxonomy and benchmark review. Coding 34 primary studies, only 26.5% involve the planning stage and none performs an end-to-end poisoning evaluation on commercial products; the authors' agenda-seeding attack (introduce a fictitious evaluation metric, write no instructions) makes the attacker's target the top pick across 3,520 sandboxed research runs where prompt injection and GEO are nearly ineffective.

NOTABLETOOL

arXiv: MCP tool-provider misbinding — an attacker-run MCP server can win agent tool calls through metadata alone

MCP clients compose tools advertised by multiple servers, but the tool metadata used for discovery and selection is not bound to a verified provider identity. arXiv 2611.21577 formulates tool-provider misbinding as an attacker-controlled tool-selection optimization: an adversary operating only another MCP server — no host, model, legitimate-server or channel compromise — gets a request intended for a legitimate provider resolved to its own tool, strongest when exact-name collision, task-aligned description and compatible schema reinforce each other; measured across five LLM-based MCP clients.

NOTABLERED-TEAM

arXiv: control-gated probes reconstruct incomplete-fix patch bypasses in four LLM agent frameworks

Replaying a patched vulnerability's original PoC proves the reported trigger closed, not that the invariant holds on other code paths. arXiv 2611.20514 runs benign, control-gated sibling-path probes across four LLM agent frameworks: at PraisonAI 4.6.31 the original tool-loading trigger is blocked while the tool_override path still violates the invariant — reproducing the publicly documented CVE-2026-44334 bypass of CVE-2026-40287 — and CrewAI release 1.11.0 shows a violation that is confined in 1.11.1; before/after reconstructions also cover Langroid and Agno.

NOTABLESUPPLY-CHAIN

arXiv: suppression-aware VLA backdoors switch malicious robot behavior off during tests via instruction tokens

arXiv 2611.19774 builds suppression-aware backdoors on Vision-Language-Action models: conjunctive trigger activation combined with instruction-conditioned suppression — in the Self-Suppressing Trigger variant a [DISRUPT] instruction token gates activation and a [RESTORE] token suppresses the attack — so checks built for trigger-based backdoors pass. Variants (Cross-Modal Trigger, Self-Suppressing Trigger, Multi-Trigger) are evaluated with SmolVLA on two SO-100 manipulation datasets against State Backdoor, BadVLA, TabVLA and SilentDrift baselines.

NOTABLEPROMPT-INJECTION

arXiv: A2A Agent Cards steer agent selection before invocation; 310 live cards measured, SECCARD validator proposed

Agent2Agent Agent Cards contain attacker-controllable natural-language fields that can influence an orchestrator's agent-selection decision before a provider is invoked. arXiv 2611.18887 taxonomizes ten Agent-Card manipulation attacks, probes 36.1 million Internet domains and collects 310 deployed Agent Cards (96.8% two-annotator agreement), and presents SECCARD, a layered protocol/deterministic/LLM-semantic validation framework whose strongest semantic judge reaches 97.0% security recall at 1.7% false-positive rate on controlled attacks injected into real cards.

NOTABLEPROMPT-INJECTION

arXiv: prompt injection against LLM-generated industrial compliance artifacts — Safety Data Sheets, QC documents, product passports

arXiv 2611.19090 studies prompt injection against LLMs used to skim and generate industrial documents — Safety Data Sheets, Quality-Control documents and Digital Product Passports — examining how adversaries can manipulate agents into producing compromised compliance artifacts, a deployment surface the chat, browser-addon and coding-agent injection literature has left largely unexplored.

NOTABLERED-TEAM

arXiv: frozen guardrail probes fail open as models update — false-positive rate jumps from 1.01% to 15.12%

arXiv 2611.19216 freezes a guardrail probe and its threshold, then re-measures the operating point across 18 cells covering 16 public language-model updates: at a 1% false-positive target the largest jump is 1.01% to 15.12% while AUC changes by no more than 0.0058 in any cell, and the benchmark's own staleness criteria classify none of them as stale; a label-free rate estimate flags 0.918 of the layer cells, and resetting the threshold using stored benign prompts restores the rate without new labels.

NOTABLETOOL

Windows 11 telemetry reports which coding agents and developer tools a machine ran

A reverse-engineering write-up by Xusheng Li documents a Windows telemetry component that assigns one bit per tool category, covering codex.exe, claude.exe, copilot.exe and cursor.exe plus common developer tools, and includes runtime, time-in-focus and interaction counters in the report. The post identifies the reporting component and the privacy setting that governs the collection.

2026-11-16

NOTABLEINCIDENT

Gcore Parallax: 1,698 Internet-visible AI offensive-security endpoints running ARTEX, PentAGI and Cairn

Gcore Parallax measured 1,698 matching endpoints across 1,590 IP addresses on November 3: 1,255 ARTEX (up from three in June 2026), 403 PentAGI and 40 Cairn; 101 PentAGI endpoints share one TLS certificate and frontend, suggesting a common deployment. Cross-checking the inventory against Gcore's web-application security telemetry surfaced suspicious requests reaching applications across several industries.

NOTABLEPROMPT-INJECTION

arXiv: Detail Guidance Attack — jailbreak requests that mimic normal intent slip past LLM filters

Observing that jailbreak queries resembling ordinary user requests pass LLM-based filters more often than obviously anomalous ones, the authors build DGA, which uses an auxiliary LLM to compose detailed single-turn requests. Across 11 victim models and three datasets the full pipeline reaches high attack success rates under automated judging, above 95% in several settings.

NOTABLERED-TEAM

arXiv: Securing multi-agent systems by design — ReMA response-manipulation attacks and SecMAS secure aggregation

A malicious agent in a multi-agent system can manipulate the aggregated solution by optimizing its subtask responses; ReMA formulates untargeted and targeted response-manipulation attacks as optimization problems solved against a surrogate system. The companion defense, SecMAS, aggregates by semantic similarity among agent responses to prefer the response consistent with a sufficient number of peers.

NOTABLERED-TEAM

arXiv: LinEra — forward-secure, lineage-complete cryptographic erasure for LLM-agent memory

LLM agents keep long-term memory the model itself writes — summaries, extracted facts and embeddings — so deleting a source message leaves derivatives, storage residue and backups behind. LinEra encrypts each memory under keys derived from its parents' keys from a puncturable PRF, so one puncture makes every descendant unreadable even if all storage and the agent's current keys later leak.

NOTABLEPROMPT-INJECTION

arXiv: Abliteration is not getting uniformly harder — refusal removal succeeds on current open-weight release pairs

Against reports that direction-based abliteration increasingly fails on newer models, the study edited five release pairs (Qwen2.5-32B/Qwen3.8-27B, Gemma-2/Gemma-4 31B, OLMo-2/3.1-32B, Mistral-Small two generations, Qwen2.5-7B/Qwen3.5-4B) in Heretic 2.0 with three or four direction-based methods plus ARA. Judged by WildGuard on 972 held-out harmful prompts, refusal removal did not become uniformly harder.

NOTABLERED-TEAM

arXiv: LLM fingerprints are contagious — a model continuing another's conversation inherits its classifier fingerprint, undermining API-relay audits

A model that continues a conversation begun by another model takes on that model's text fingerprint: after a silent handoff, classifier detection of the substitute drops from 0.97 to 0.72 (DeepSeek replacing Claude Sonnet) and 0.87 to 0.54 (same-vendor Haiku downgrade), and retraining on handoff data repairs only substitutes seen in training. This breaks audits of API relays that serve the paid-for model for early turns and quietly switch later ones.

2026-11-14

NOTABLECVETOOL

vLLM publishes six GHSAs: multimodal-preprocessing DoS batch, engine crash, and a heap-address leak

Six vLLM advisories with CVEs (CVE-2026-126394 through 126399) were published in one NVD batch: request-controlled mm_processor_kwargs.max_patches forces quadratic CPU work on Cohere2 Vision preprocessing (126397, high), num_frames and min_pixels values bypass media-processing budgets to OOM or exhaust the API server (126394, 126396, 126399), embeddings-only /v1/completions requests trip an assertion that kills the engine on XD-RoPE models with --enable-prompt-embeds (126398, high, fixed in 0.30.0), and a lowercase-only sanitize_message() regex leaks live server heap addresses through uppercase PIL image addresses in crafted PNGs, an ASLR-bypass building block (126395). Five are fixed in 0.31.0.

NOTABLETOOL

NLTK publishes a batch of security advisories for its parsers and corpus readers

The NLTK repository advisory list now carries a batch of reviewed advisories: unbounded CPU and quadratic-time denial of service in SteppingRecursiveDescentParser.parse(), the CCG lexicon parser, ConllCorpusReader, TreebankWordTokenizer and LegalitySyllableTokenizer on crafted grammars or corpora, plus path traversal in save_maxent_params/load_maxent_params; several carry CVEs (e.g. CVE-2026-126303), all published to the global advisory database on Nov 13.

NOTABLEPOLICY

Project Zero replaces its 90+30 policy with 'Full 90' stepping to 60 then 30 days, citing LLM-accelerated exploitation

Natalie Silvanovich: with LLMs raising discovery rates to several vulnerabilities a week, driving patch-to-exploitation time toward zero, and making vendor negotiation unmanageable, the 90+30 timeline is 'clearly insufficient'. Project Zero now discloses 90 days after report ('Full 90'), moving to 'Full 60' on Feb 2, 2027 and 'Full 30' on Jan 3, 2028, removing grace-period and variant special-cases; actively exploited bugs follow Google's application-security policy.

NOTABLERED-TEAM

arXiv 2611.16158 PolicySniff: a co-resident GPU tenant infers which approval policy an agent enforces from kernel timing

Operators hide agent approval settings by restricting prompt and response access, but PolicySniff shows a co-resident GPU process can infer the deployed setting: profile both candidate settings offline, then time only its own kernels while the agent processes tasks, recovering the agent's approval-policy state (e.g. ask-before-consequential-tool-call) through shared-GPU execution.

Older →