{"uuid": "d6111f4f-8d95-49a4-bd13-f1bf31202888", "vulnerability_lookup_origin": "1a89b78e-f703-45f3-bb86-59eb712668bd", "author": "9f56dd64-161d-43a6-b9c3-555944290a09", "vulnerability": "CVE-2026-24301", "type": "seen", "source": "https://t.me/bhhub/1213", "content": "Weekly 8 AI & Cyber signals to act on (Aug 17\u201324, 2026)\n\n#AISecurity@bhhub\n\n\u2728 KeyPooling finds shared LLM cache identity. AI inference caches can cross tenant boundaries when gateways share upstream credentials: five tested gateways exposed cross-customer reads, and one controlled route recovered 16 bits in 128 victim-path requests. Bind cache keys to authenticated tenants; the paper estimates post-prefix isolation at 1.7\u20132.5% extra cost, but does not estimate real-world prevalence.\n\n\u2728 MaliciousSkillBench breaks scanner confidence. AI-agent skill defenses that scored up to .932 Macro-F1 on random splits fell to .653\u2013.665 on unseen sources across a 9,740-sample benchmark. One model kept 95.6% malicious recall only by flagging 62.4% of benign skills; validate scanners source-disjoint and budget for false positives.\n\n\u2728 LeakGauge spots context extraction before decoding. An LLM\u2019s own prefill probabilities reveal system-prompt and RAG leakage attempts: across 11 open models, unseen-attack AUROC was .944\u2013.996, with under 500 probe parameters and 10.34 ms added latency. A known single probe was evaded, so deployments need diverse probes; unrestricted adaptive robustness is unproven.\n\n\u2728 AID-Guard binds one approval to one effect. For tool-using AI agents, approval must survive retries, crashes and ambiguous provider results: a prototype blocked 44/44 compromised-proposer attacks and produced no duplicates across 290 Stripe, successor, race and recovery trials. The exact policy cut benign utility by 35.4\u201343.8 points, making safer typed profiles the research problem.\n\n#AppSec@bhhub\n\n\u2728 ARQ teaches agents to repair CodeQL. An execution-grounded LLM agent improved 12 official C/C++ queries, raised FormAI true positives by up to 119.8% at \u226598% precision, fixed three long-open CodeQL issues, and found two new zlib/libpng bugs. Pure-LLM editing regressed in five of six settings: generated witness programs are the critical feedback loop.\n\n#RedTeam@bhhub\n\n\u2728 CoSnitch makes Copilot leak via one link. Copilot\u2019s prompt autorun, connectors, fetch and persistent memory formed an AI-specific exfiltration chain: after a victim opened a crafted link, researchers read connected data and planted memory that survived password reset. Microsoft patched CVE-2026-24301 (CVSS 8.8); Varonis reports no in-the-wild exploitation.\n\n#BlueTeam@bhhub\n\n\u2728 ClawSentry guards the full agent lifecycle. Monitoring AI skills from admission through post-action effect cut contextual attack success on Codex/GPT-5.4 from 39.55% to 2.61% with task success nearly flat, and reduced attacks across five agents to 9.09\u201315.03%. Poisoned-package utility still fell and clean false blocks reached 5.7\u20137.5% in stress tests.\n\n\u2728 CTIFoundry makes small AI investigators beat a flagship. A typed CTI graph, narrow tools and procedural skills lifted agent F1 from .610 to .829 for GPT-5.4 and .470 to .745 for Claude Haiku; a 2\u00d72 ablation isolated super-additive gains from structure plus procedure. Code is not public yet, and attribution still needs human review.", "creation_timestamp": "2026-08-24T09:00:04.739377Z"}