Apollo embedded evaluators bound by OpenAI whitelist: no publishing without approval, redactions allowed

BlackHC · x · 2026-11-17

Daniel Kokotajlo revealed that Apollo's embedded evaluators may only talk to a whitelist of OpenAI-chosen employees, must notify OpenAI of each conversation, cannot publish results without OpenAI approval—and OpenAI can redact anything or even withhold in-scope information entirely on security/IP/"technical" grounds. He calls the arrangement "extremely weaksauce."

BlackHC comments that this makes the evaluators sound "employee-like after all," raising questions about the independence of external safety oversight.

Related event: OpenAI Reportedly Restricts Apollo Evaluators with Whitelists and Gag Terms(2 posts)→

Original post →

More from Safety

Safety channel →