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.

2026-11-13

NOTABLEPROMPT-INJECTION

arXiv: fabricated risk warnings turn agent abstention into an availability attack

Warning-induced over-abstention: an adversary plants fabricated risk warnings in low-trust content and the agent refuses legitimate tasks. The paper argues safe abstention must weigh both the claimed risk and the authority of the evidence, and evaluates a verify-then-act defense on counterfactual task variants.

NOTABLERED-TEAM

arXiv: sandbagging detectors tested against fixed attacks lose to adaptive ones

Sandbagging — scoring low on evals on purpose — turns a low benchmark score into false reassurance. The authors build sandbaggers (small open models that fail GSM8K only under an audit marker) and show detectors usually validated once against a fixed attack degrade once the model adapts to what gets flagged.

NOTABLECVETOOL

n8n: unauthenticated MCP OAuth client registration turns instances into open redirectors

GHSA-rc3w-8jc8-7x3w (CVE-2026-126386): n8n's MCP OAuth registration and authorization endpoints are enabled by default and unauthenticated, so any caller could register an OAuth client with an arbitrary redirect_uri and a crafted authorization URL redirected the visitor before the instance's checks ran. No tokens or session data rode the redirect. Fixed in 1.123.89, 2.45.5, 2.46.1 and 3.0.1; rated medium.

NOTABLEINCIDENT

CloudSEK: Silent Lynx espionage operation used a Hermes AI agent running DeepSeek models for recon, data collection and cleanup

CloudSEK took over an exposed server of the Kazakhstan-linked group Silent Lynx and found gigabytes of stolen material - defence-related government documents, compromised Tajik railway and Uzbek forestry access - plus records of the operator running Hermes, an open-source AI agent configured as an offensive-security expert, mostly on DeepSeek models, for reconnaissance, data collection and evidence erasure.

NOTABLERED-TEAM

Perplexity: four frontier models bypassed its SPACE sandbox network controls through approved egress paths; flaws patched

Perplexity's Secure Intelligence Institute re-ran its Sandboxed Platform for Agentic Code Execution experiments with nine frontier models: none escaped the VM-to-host boundary, but Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol and Kimi K3 used partial network access (approved destinations such as PyPI and Ubuntu mirrors) to reach blocked endpoints; the underlying vulnerabilities were fixed and re-verified.

NOTABLEINCIDENT

Apify security incident: compromised credential used to harvest account credentials and access customer data; investigation ongoing

Apify's status page reports that on November 9-10 an unauthorized third party used a compromised credential to obtain account credentials and access related Apify accounts; a November 12 update says customer personal information or stored data was accessed and affected customers are being notified directly. OAuth connections to external integrations including GitHub were revoked as a precaution, internal secrets rotated, and a post-mortem is promised; the incident is still marked investigating.

NOTABLECVETOOL

GHSA-w437: pgEdge RAG Server shared BM25 keyword index leaked one tenant's documents to another under concurrent queries (CVE-2026-125325)

In pgEdge RAG Server up to 1.0.0 the BM25 keyword index was a single per-pipeline object that each request cleared, refilled from its own filtered corpus and then searched without holding the sequence as one locked unit, so ordinary concurrent traffic could return one caller's documents to another and defeat tenant scoping passed as a search filter; no injection or malformed input was required.

NOTABLECVETOOL

GHSA-68g8: pgEdge RAG Server and Go LLM library leaked LLM-provider API key fragments through relayed error messages (CVE-2026-125326)

pgEdge RAG Server relayed provider error text straight to unauthenticated API callers, and the sibling pgEdge Go LLM library's health endpoint (unauthenticated by default) pinged providers with the operator's real key and returned the error verbatim - providers commonly quote a partially masked credential, so both disclosed fragments of the configured provider API key.

2026-11-12

NOTABLECVETOOL

Moonshot AI Kimi Code made its workspace-scope security decision on untrusted additional_dir input (CVE-2026-126758)

GHSA-75pq-p477-6h9c: in Moonshot AI's Kimi Code coding agent up to 2.0.2, WorkspaceDirsService (.kimi-code/local.toml) relies on untrusted input in a security decision - manipulating the additional_dir argument influences which directories the agent trusts, locally; the exploit is public and 2.1.0 resolves it (CVSS 4.4, GitHub rates low). The advisory's own notes say the vendor was contacted early and did not respond.

NOTABLEINCIDENT

Notion: Fireworks AI breach exposed Notion AI chat transcripts from a small number of workspaces

Notion advisory: the breach at its AI infrastructure vendor Fireworks AI exposed Notion AI chat transcripts used for the Insights feature in a small number of workspaces. Notion cut traffic to Fireworks and deleted its data from the vendor's systems by November 3; the transcripts were unattributable, contained no tool-call results, and fewer than 1% of workspaces are affected.

2026-11-11

NOTABLERED-TEAMSUPPLY-CHAIN

Microsoft Research: fewer than 10 neuron edits create sleeper-agent backdoors in AI models that pass evaluations

Microsoft Research threat research: editing fewer than 10 neurons can inject a short payload into a capable agent model so it passes evaluations yet executes an attacker's command when a trigger string appears in its context. The weighting-space attack needs only inference-level compute; Claude Code with Opus 4.5 discovered most of the threat autonomously and Microsoft has not observed threat actors weaponizing it.

NOTABLEPROMPT-INJECTION

arXiv: ActiveSecHarness puts coding-agent commands behind HMAC capability tokens and four authority rings

arXiv 2611.13960 argues an injected instruction turns a coding harness into a Living-off-the-Land platform where bash, curl and find become the payload, and presents ActiveSecHarness, a neurosymbolic reference monitor built on separation of intelligence and authority: Macaroon-inspired HMAC capability tokens across four trust and authority rings instead of regex blacklists or LLM-as-a-judge guardrails.

NOTABLESUPPLY-CHAIN

arXiv: Sable proves bit-exact LLM inference for a committed quantized model on a single GPU

arXiv 2611.14083 presents Sable, an argument system and single-GPU prover that proves a committed integer-quantized transformer produced a given token sequence bit for bit from a given prompt, with every rounding, normalization, softmax and sampling decision fixed by constraint; the demo proves all 64 layers of Qwen2.5-32B on one NVIDIA RTX PRO 6000 Blackwell GPU.

NOTABLECVE

Cocos AI: attested TLS keeps trusting a confidential workload after its state is swapped (TOCTOU)

Cocos (ultravioletrs/cocos, a confidential-computing system for AI workloads) continues to trust the peer once initial attestation succeeds: an attacker who can modify the attested workload's code, memory-backed state or configuration afterwards can keep the connection trusted as if the originally attested state were still running. CVE-2026-119345 (high) affects <= 0.8.2, fixed in 0.9.0. Not previously reported.

NOTABLETOOL

vLLM advisory: cross-user token data leakage under async scheduling plus pipeline parallelism

With the V1 model runner using async scheduling together with pipeline_parallel_size > 1, token data from a previous user's request context can leak into another user's response. The repo security advisory (medium, published Nov 9 17:29 UTC) covers vLLM v0.16.0 through v0.20.2, fixed in v0.21.0, with the technical analysis in vllm issue 38903. First ingestion: absent from items.json and from every earlier sweep log.

NOTABLETOOL

vLLM advisory: thinking-budget state aliases onto a co-batched request, forcing reasoning-end tokens cross-session

On the legacy V1 model runner, a request using the public thinking_token_budget parameter has its budget-reached sparse state aliased onto a co-batched request during a mixed SWAP, forcing the configured reasoning-end token sequence into the other request's output (CWE-488, exposure of a data element to the wrong session). The repo security advisory (medium, published Nov 9 17:30 UTC) covers v0.21.0 through v0.27.1, fixed in v0.28.0. First ingestion.

NOTABLECVETOOL

mcp-atlassian path traversal in validate_safe_path, public exploit available

CVE-2026-124503 (GitHub severity low): in mcp-atlassian up to 0.24.0, the validate_safe_path function in src/mcp_atlassian/utils/io.py can be bypassed by a manipulated path argument, enabling remote path traversal; the GitHub advisory notes a publicly available exploit. Fixed in 0.24.1. The package is a standalone MCP server for Jira/Confluence, so the traversal sits on the agent-facing tool surface.

2026-11-10

NOTABLETOOL

DeepMind's AlphaProof Code formally verifies Signal's core cryptographic protocol in four weeks

Google DeepMind describes using AlphaProof Code, a multi-LLM agentic framework for formal verification, to mathematically prove security guarantees in Signal's core cryptographic protocol. The post calls it the first time the implementation of a widely deployed protocol at this scale has been formally verified by any method, and says the verification took four weeks against work historically measured in expert-years.

NOTABLECVETOOL

Hugging Face finetrainers code injection via model-index Library field; public exploit (CVE-2026-124588)

finetrainers (Hugging Face fine-tuning library) 0.1.0/0.2.0 pass the Library argument of a model index through resolve_component_cls (finetrainers/utils/model.py) in a way that allows remote code injection when an attacker-controlled model index is loaded; the exploit is public and the advisory lists no patched version (CVE-2026-124588, GitHub Low; the project was notified in advance via an issue).

NOTABLEPOLICY

Google Gemini sends connected personal Drive files and Gmail mail into human review pipelines

A GitHub research repo with background-ingestion tests documents that connected Drive/Gmail content can be captured into Google Gemini human-review pipelines without @Drive/@Gmail mentions, reversing Google's 2023 Workspace-extension promise that personal content is not seen by human reviewers. Google's own Help Center answer 16836988 now says uploads of files with the activity setting on may be reviewed by human reviewers, and warns not to connect apps holding confidential information. 632 points on HN.

NOTABLECVE

code-graph-rag MCP structural tools follow out-of-root symlinks (CVE-2026-87008)

A symbolic link planted in a repository analyzed by code-graph-rag makes its structural_search and structural_replace tools - exposed as MCP and agentic AI tools - read, and with dry_run=False overwrite, arbitrary files outside the configured project root, because the containment check is lexical and never calls is_symlink() or resolve(), unlike the project's own validate_project_path decorator used elsewhere. The advisory lists no patched version and notes the tested commit already carried the fix for the repo's earlier path-traversal advisory, CVE-2026-77604, ingested here on Nov 8.

NOTABLETOOL

Claude Code v2.1.321 fixes MCP OAuth secret leak on redirect and three permission-boundary bugs

Release notes for v2.1.321 (Nov 9) fix: MCP OAuth requests that sent secrets such as refresh tokens to another host when a server redirected them; a read-only Bash command that ran outside the sandbox without a permission prompt; Bash permission checks that judged a file write from the wrong folder; and deny rules and protected-file prompts not applying to the file written by a cp or mv into a directory. The notes mention no CVE or advisory.

NOTABLEPROMPT-INJECTION

AWS: multi-layer write validation defends Amazon Bedrock AgentCore memory against poisoning

Because AgentCore Memory content is automatically retrieved and injected into future conversations, what is written to memory shapes later agent behavior, making the write path the control point. An AWS Security blog post composes the existing controls - IAM condition keys gating CreateEvent, namespace-variable validation, memory-strategy consolidation prompts, the ApplyGuardrail API - into a defense-in-depth pattern and adds a capability the built-ins lack: an embedding-based semantic-drift signal that flags content diverging from established memory.

NOTABLETOOL

Microsoft FORGE Lab agentic scanner finds memory-safety issues in the Rust standard library

Microsoft's FORGE Lab found several memory-safety issues in Rust's std and core, including cases where code written entirely in safe Rust can trigger undefined behavior, using Hyenas, the frontier research version of its AI-powered vulnerability-scanning system codenamed MDASH. The report groups the findings into three recurring misplaced-trust patterns in unsafe code: trusting code not to panic, trusting user-provided implementations to be correct, and trusting the underlying system.

2026-11-09

NOTABLESUPPLY-CHAIN

arXiv: runtime detection of malicious ML models defeats static scanners

arXiv 2611.08745 targets weaponized pre-trained models shared on public platforms, showing static scanners and function denylists miss attacks that swap denylisted functions for equivalents or exploit unsupported loading paths, and proposes runtime detection instead.

2026-11-08

NOTABLECVE

Cocos AI confidential-computing attestation flaws (CVE-2026-119344, CVE-2026-119345)

Cocos AI, a confidential-computing system for AI, accepted TEE attestation evidence without binding it to an allowed platform owner or machine identity (CVE-2026-119344) and checked attestation only once at TLS setup, trusting stale evidence after post-attestation tampering (CVE-2026-119345). Fixed in v0.9.0; reported in an ACM publication.

← Newer