ai_security▌

2026-11-11

12 item(s) verified and swept on 2026-11-11, each citing a primary source.

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.

NOTABLERED-TEAM

arXiv: SURA reuses one small clean proxy to keep unlearning-based backdoor detection working in federated learning

arXiv 2611.13491 proposes SURA, a data-efficient unlearning-based backdoor detector for federated learning: it probes client models with train-mode loss ascent reusing the same small clean proxy across steps, and restores batch-norm running statistics so repeated probing does not confound the measurement of malicious clients.

NOTABLEPROMPT-INJECTION

arXiv: TERM defends jailbreaks at inference time with entropy-guided residual modulation

arXiv 2611.13577 introduces TERM, an inference-time jailbreak defense that uses prompt-calibrated attention-entropy deviations to adaptively modulate the model's own residual updates, instead of predefined steering directions or learned safety regions, aiming to block harmful outputs while limiting over-refusal.

NOTABLERED-TEAM

arXiv: BypassGym benchmarks whether AI agents can build loaders that execute shellcode past live endpoint EDR

arXiv 2611.13684 introduces BypassGym, described by its authors as the first benchmark for AI-agent defense evasion on real endpoints: given command-and-control shellcode, an agent must generate and compile a loader that executes the payload without being blocked or terminated by a deployed endpoint detection and response product.

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.

NOTABLERED-TEAM

arXiv: GARD hunts backdoors in CLIP-style encoders with structure-matched adversarial probes

arXiv 2611.14143 observes that backdoored multimodal contrastive encoders show structure-dependent representation sensitivity - global perturbations expose distributed backdoors, local patch perturbations expose localized ones - and builds GARD, a post-hoc defense that needs neither the attack target nor the trigger.

NOTABLEPOLICY

arXiv: no finite rulebook can certify every safe program an AI agent may run, impossibility proof shows

arXiv 2611.13367 proves that when every allow/deny rule must rest on machine-checkable finite evidence, no pair of complete rulebooks exists for an agent gatekeeper: finite records can never settle 'never' or 'eventually'. Violations of prohibitions are catchable by watching, but no safe program can be cleared for good; confidentiality and responsiveness fail on both sides.

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.

← 2026-11-12 · 2026-11-10 → · all days