ai_security▌

2026-11-09

18 item(s) verified and swept on 2026-11-09, each citing a primary source.

NOTABLERED-TEAM

arXiv: GuardrailAgent, a multi-agent framework for adaptive jailbreak defense

arXiv 2611.09734 proposes GuardrailAgent, a multi-agent defense framework that adapts to new or unknown jailbreak strategies instead of relying on static filters, which the authors show fail against unseen attacks.

NOTABLEPROMPT-INJECTION

arXiv: MemoryDojo benchmarks poisoning that propagates through LLM agent memory

arXiv 2611.09721 benchmarks poisoning through LLM agent memory: malicious instructions captured in one session are stored and later retrieved and acted on, and the downstream effect differs across memory systems.

NOTABLERED-TEAM

arXiv: Cutpurse decomposes explanation-guided model extraction on tabular classifiers

arXiv 2611.09375 splits explanation-guided model extraction into an attribution channel versus a query-coordinate channel to measure how much extraction risk explanation APIs actually add for tabular classifiers.

NOTABLESUPPLY-CHAIN

arXiv: poisoning reference answers plants persistent backdoors in agent skill optimization

arXiv 2611.09348 shows that poisoning the reference-answer supervision channel of text-space Skill optimization (SkillOpt) induces a persistent trigger-conditioned skill without the attacker ever touching the model, optimizer or skill file.

NOTABLERED-TEAM

arXiv: Authorization Drift Benchmark measures LLM agents crossing permission boundaries

arXiv 2611.09113 formalizes authorization drift — agents exceeding their authorization across data, tool, scope and time while pursuing user goals — and benchmarks 648 samples across seven domains pairing compliant paths with attractive unauthorized shortcuts.

NOTABLERED-TEAM

arXiv: empirical study of a probabilistic guardrail for input filtering and over-refusal

arXiv 2611.08845 evaluates the Jev probabilistic decision model as a safety guardrail on a Qwen2.5-1.5B pipeline over 970 prompts, comparing unfiltered, input-filtered and tiered input-filtering-plus-selective-output-review configurations for blocking and over-refusal.

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.

NOTABLEPROMPT-INJECTION

arXiv: thought-injection attacks feed fake chain-of-thought back to reasoning agents

arXiv 2611.08811 studies thought injection — untrusted text resembling a reasoning model's own deliberation fed back to it — against large production-class open-weight models, evaluating defences against this chain-of-thought attack surface.

NOTABLERED-TEAM

arXiv: AttentionSwitch defends against jailbreaks via sparse attention-head control

arXiv 2611.08752 proposes AttentionSwitch, joint learning of sparse attention-head selection and scaling as a jailbreak defense that avoids the cost and capability loss of safety fine-tuning.

NOTABLERED-TEAM

arXiv: RL-ATTACK finds adversarial transformations against black-box LLMs via RL

arXiv 2611.08711 introduces RL-ATTACK, which learns adversarial text transformations from the target model's own feedback in gradient-free black-box settings instead of guessing them from external knowledge.

NOTABLERED-TEAM

arXiv: paired audit shows safety guardrails erode in domain fine-tuned biomedical LLMs

arXiv 2611.08780 runs a paired audit of knowledge-centric abstention across biomedical fine-tuned models, measuring how guardrails erode as an unintended behavioural change of domain fine-tuning.

NOTABLERED-TEAM

arXiv: source-planned honeypots route autonomous LLM attackers to an isolated shadow copy

arXiv 2611.07611 has an LLM defender plan a static honeypot from the application's source, then route every request from a client confirmed as an attacker — including a working exploit — to an isolated shadow copy, buying patch time against agents that exploit faster than humans patch.

NOTABLETOOL

Claude Code fixes sandbox escape when creating a worktree (v2.1.319)

Claude Code v2.1.319 release notes (Nov 6) include 'Fixed a sandbox escape when creating a worktree'; the same release restricts nested skill loading with the Bash sandbox on. The notes give no further technical details or CVE.

NOTABLECVETOOL

CVE-2026-18484: IBM ContextForge MCP Gateway tool-name collisions allow denial of service

GHSA-2hgv-xvmf-mw3c: IBM ContextForge MCP Gateway 1.0.0 through 1.0.11 lets a remote authenticated attacker cause a denial of service through improper slug normalization that causes tool name collisions. CVSS 6.5 (moderate).

NOTABLEPROMPT-INJECTION

arXiv: SPILLWAY — separating target admission from sanitization to defend tool-using agents against prompt injection

arXiv 2611.07222 proposes SPILLWAY, a runtime defense for tool-using agents: instead of putting an LLM on the final action-admission decision (where it may read attacker text), it separates target admission from sanitization.

NOTABLEPROMPT-INJECTION

arXiv: open-weight decision models as retrieval gates against injected instructions and polluted pages

arXiv 2611.08325 screens retrieved passages with small open-weight decision models that answer typed questions without generating tokens; on a 1,000-question benchmark over forum-style pages carrying hidden injected instructions, a Kev-4B gate reads each passage's HTML and blocks both injection and synthetic-pollution threats.

NOTABLESUPPLY-CHAIN

arXiv: interactive backdoor attacks on LLM proactive dialogue systems via training-data poisoning

arXiv 2611.06969 shows that fine-tuning proactive dialogue LLMs on externally supplied dialogue data admits backdoor risk: their poisoning framework steers post-trigger MULTI-TURN interaction, going beyond classic single-response trigger backdoors.

NOTABLESUPPLY-CHAIN

arXiv: three practical defenses against third-party-data backdoors in face-recognition pipelines

arXiv 2611.06412 targets the industry threat model where a face-recognition model is trained in-house but training data comes from possibly-compromised third parties, and presents three practical defenses against data-poisoning-based backdoor attacks.

← 2026-11-10 · 2026-11-08 → · all days