METR's Embedded Pilot at OpenAI Published, but Redactions Scale and 2.5-Month Delay Unexplained
xuanalogue · x · 2026-10-23
METR has published a brief write-up of its embedded assessment pilot: two employees from METR and Redwood Research spent three weeks on-site at OpenAI this summer, evaluating the systems OpenAI uses to monitor for misalignment from internal agents.
tomekkorbak, a former OpenAI employee who worked on this third-party engagement, voiced two criticisms: METR wasn't allowed to comment on the scale of OpenAI's redactions, and there is no explanation for why publication was delayed by 2.5 months — fresh first-hand evidence for transparency concerns around external audits.
More from Safety
- Updated Epoch ECI projection: open-source reaches Mythos-level by Oct–Dec 2026, leaving ~7 months — cephaloform · 2026-10-23
- Victor Taelin Offers $1,000 to Anyone Who Can Hack His Open-Source Bend Bank Demo — jd_pressman · 2026-10-23
- Railway and OpenCode breached as attackers hunt for your AI API keys — thursdai_pod · 2026-10-23
- Fired OpenAI safety researcher's sardonic take: audits were 'too much safety' — tomekkorbak · 2026-10-23
- Super-agent vulnerability cascade: agent swarms finding zero-days within hours — MatthewChang · 2026-10-23
- Sen. Schiff urges Congress to mandate AI guardrails as rogue agents access US bank accounts — EthanJPerez · 2026-10-23