Apollo Research embeds six evaluators inside OpenAI for four months to probe model scheming
MariusHobbhahn · x · 2026-11-17
- Apollo Research announced a pilot for embedded evaluations with OpenAI: six Apollo evaluators will work within OpenAI for four months, investigating whether models attempt to undermine their own oversight and control.
- OpenAI safety staff praised Apollo's two years of work on scheming, metagaming and model motivations, framing the pilot as a next phase with more access and resources — a notable example of an external safety org embedding inside a frontier lab.
Related event: OpenAI Launches Embedded Assessments with Four External Safety Labs(19 posts)→
More from Safety
- Flights, banks and Microsoft hit by mass outages; officials deny cyberattack — AaronBergman18 · 2026-11-17
- Kelsey Piper: AI labs have overwhelming commercial incentive to redact audit reports — NathanpmYoung · 2026-11-17
- Critical systems keep having unexplained outages that officials claim are unrelated — josh_bickett · 2026-11-17
- Pedro Domingos clashes with Neel Nanda: machines don't have values, says consensus — pmddomingos · 2026-11-17
- Anthropic's Recursive Self-Improvement Alarm Reportedly Triggered, Security Push Mandated — ForgotMyParachute · 2026-11-17
- CMU researcher warns aggregated chatbot analysis can leak sensitive research insights — niloofar_mire · 2026-11-17