$ cat briefs/daily/2026-10-09.md
2026-10-09
October 9, 2026 (Fri)
3 stories · 2 paper picks · 2 new pages
Top Stories
1. Anthropic moves cyber defense toward patch delivery
The Cyber Mission combines an infrastructure-defense program with OSS Scanner, free periodic scans for participating open-source projects. Scanner reports arrive without human review; Anthropic expects above 90% true positives, rather than reporting an independently measured result.
Why it matters: maintainers gain faster findings, but still need capacity to verify and prioritize them. → AI-Enabled Cyberattacks (source)
2. Claude's revised policy adds physical-action controls
Effective 2026-11-12, the policy requires qualified operators to observe and stop potentially injurious autonomous equipment, and equipment to remain safe if Claude disconnects. It also removes the blanket ban on personalized campaign targeting while retaining prohibitions on deception and privacy abuse.
Why it matters: deployment rules now explicitly cover the transition from recommendations to physical action. → MHS — Model Hardware Standard · AI Governance (source)
3. OpenAI reports its first disrupted Category 5 influence operation
OpenAI rates a Russia-origin operation Category 5 and an Iran-origin operation Category 4 on the 1–6 Breakout Scale. Both reached established media; their own internal impact reports exaggerated effectiveness. These are reach assessments, not measured persuasion effects.
Why it matters: publisher identity and distribution channels matter alongside whether content was AI-generated. → Content Provenance (AI output marking) (source)
Paper Picks
RunningTab — remember what the deliverable still owes. An environment-side record connects requirements to read excerpts and unopened files, then checks unresolved requirements before completion. Read it for a concrete way to prevent omissions; the abstract reports gains across three benchmarks and three LLMs without numerical scores. → RunningTab — tracking unfinished workspace requirements (source)
MIMESIS — train with less accommodating simulated users. A 9B user simulator learns 13 behavioral patterns. Agents trained with it outperform GPT-5.5-trained agents under all nine unseen user simulators across eight environments. Read it for training-environment design; this does not establish transfer to real users. → MIMESIS — training agents with learned user behavior (source)
Watch
- Negative-Policy OPD: weaker-policy rollouts add corrective training examples while preserving teacher supervision. The abstract claims gains across settings but gives no numerical improvement or compute comparison. → Agentic Reinforcement Learning (source)
New in Wiki
- RunningTab — tracking unfinished workspace requirements and MIMESIS — training agents with learned user behavior — research notes with methods and evidence limits. (workspace record) (user simulation)