{"uuid": "dda4a57a-f7d3-4aac-a45b-61be764e2a75", "vulnerability_lookup_origin": "1a89b78e-f703-45f3-bb86-59eb712668bd", "author": "9f56dd64-161d-43a6-b9c3-555944290a09", "vulnerability": "CVE-2026-53413", "type": "seen", "source": "https://t.me/bhhub/1211", "content": "Weekly 8 AI & Cyber signals to act on (Aug 10\u201317, 2026)\n\n#AISecurity@bhhub\n\n\u2728 Encrypted reasoning traces leak across LLM sessions\nClient-returned AI reasoning is a cross-user attack surface: researchers moved encrypted blocks into a weaker model from the same provider and recovered plaintext across Anthropic, OpenAI, and Google. Decoding 315,320 public-repository blocks exposed 367 PII artifacts and 182 credentials, plus hidden hazardous reasoning. Bind traces to user, session, and model; this is a preprint.\n\n\u2728 PIPES gives agent observations provenance\nAI agents can mistake attacker-controlled tool fields for authoritative state, making unsafe actions look justified. PIPES checks response units against semantic priors and provenance; on six splits it cut Gemma 4 31B attack success from 84.7% to 2.3% while benign utility rose from 90.6% to 92.5%. The preprint tests two agents and assumes trusted metadata.\n\n\u2728 ColluSkill hides one attack across three skills\nLLM planning turns plausible agent skills into one malicious workflow: ColluSkill splits intent across packages, then refines them against scanners. On 200 three-skill chains and six scanners, attack success reached 96.0%; ChainGuard reduced it to 22.5% while passing 99.5% of benign workflows. Scan capability and artifact flow across the installed set.\n\n#AppSec@bhhub\n\n\u2728 PySeType finds nine confirmed zero-days\nLLMs supply semantic types that static analysis cannot infer from syntax: PySeType asks models to check natural-language security meanings, then traverses Python control flow with pytype. Across 103 web projects it reports 87% precision, 88% accuracy, and 15 potential zero-days, nine developer-confirmed. Dynamic loading, callbacks, and cross-language flows remain blind spots.\n\n\u2728 SRE-Bench tests agents on unseen binaries\nAI security capability drops when source code becomes a protected binary: SRE-Bench contributes 19 private programs, 44 in-house anti-analysis primitives, 262 instances, and 1,572 deterministic tasks built over 5,000 expert hours. The best of five frontier models scored 61.4% per instance but fully solved only 31.5%. Binary reverse engineering needs contamination control, realistic scale, and protection-aware evals.\n\n#RedTeam@bhhub\n\n\u2728 ZOOMSDAY: an AI agent finds a zero-click RCE chain\nAn AI coding agent materially accelerated closed-source exploit research: A Security reports finding and exploiting Zoom\u2019s annotation parser in under 24 hours and fewer than 20 prompts. A crafted 745-byte packet reaches a 128-byte stack buffer via an unchecked count; Zoom corroborates CVE-2026-53413/414/415 and patched affected clients. The RCE demo is macOS and the AI-speed claim is first-party, but shared code makes patching urgent.\n\n\u2728 MazeRunner branches through black-box pentests\nThree LLM agents preserve clues, revise failed actions, and switch attack branches instead of following one linear pentest plan. On ten recent HTB targets under a 20M-token cap, MazeRunner completed 47.7% of subtasks versus 36.2% for PentestGPT-V2, reached user access on six targets, and root on two; both baselines reached user on two and root on none. Human approval is still needed for destructive live actions.\n\n#BlueTeam@bhhub\n\n\u2728 GraphRAG makes threat hunts survive IOC rotation\nKnowledge-graph retrieval pushes AI-generated hunt plans beyond disposable indicators toward behaviors and techniques. In an APT28 case, GraphRAG preserved 100% of detections after every IP, domain, and hash was rotated; naive RAG preserved 29%. Nine reports from four vendors showed the same direction, but prompt wording mattered nearly as much and the corpus is small. Evaluate detection durability, not rule count.", "creation_timestamp": "2026-08-17T08:00:03.938053Z"}