Apollo Research embeds six evaluators inside OpenAI to probe model scheming
MariusHobbhahn · x · 2026-11-17
Apollo Research and OpenAI are launching a pilot for embedded evaluations: six Apollo evaluators will work inside OpenAI for four months, investigating whether models attempt to undermine their own oversight and control, with a focus on how the training stack might incentivize scheming and subversion. CEO Marius Hobbhahn calls it the culmination of two years of the lab's research.
More from Safety
- Apollo reveals limited OpenAI evaluation access, casting doubt on Altman's 'employee-like access' pledge — sjgadler · 2026-11-17
- OpenAI Explored $5B-Plus Acquisition of Interpretability Startup Goodfire — kimmonismus · 2026-11-17
- Cameron: Open weight models are America's bulwark against centralized, dystopian AI — markjeffrey · 2026-11-17
- NYT turns to Vector's Shaina Raza: election AI threat shifts from deepfakes to viral memes — VectorInst · 2026-11-17
- AI Now's Amba Kak joins Fortune Innovation Forum panel on reining in AI companies — AINowInstitute · 2026-11-17
- United Bans Unauthorized AI Agents From Booking Flights, Southwest Embraces ChatGPT — Future_AGI · 2026-11-17