OpenAI Embeds Apollo, Goodfire, METR, Transluce to Probe AI Subversion
Summary
OpenAI embeds safety evaluators from Apollo Research, Goodfire, METR, and Transluce as a pilot, with Apollo's starting this week to test whether models subvert their own or successors' alignment controls. Apollo and Goodfire are paid for the work; METR and Transluce are not, though they may accept compute or API credits.
Key Points
- OpenAI fired three researchers for allegedly sharing confidential information with METR, the firm that ran the Hugging Face incident post-mortem.
- Transluce plans to test in 2027 how OpenAI uses its safety evaluations to decide whether a model is safe enough to release.
- Redwood CEO Buck Shlegeris says the agreements fall short of employee-like access, calling that disappointing but not surprising.