DevelopingSafety
OpenAI embeds outside evaluators to test whether its models subvert oversight
Apollo Research says a six-person team will work inside OpenAI for four months to test whether models subvert their own or successors' alignment and control mechanisms. Apollo plans two public reports, and says OpenAI may redact confidential evidence but not verdicts. OpenAI also lists METR, Transluce and Goodfire as assessors.
- HEAT
- OUTLETS
- 4
- NEW IN 6H
- +3
- HACKER NEWS
- 24 pts
- FIRST SEEN
- LAST SIGNAL
Primary source
Apollo Researchapolloresearch.ai/blog/pilot-towards-embedded-evaluationsCoverage · 6 articles, oldest first
Monday, November 16, 2026
Apollo Research
apolloresearch.ai/blog/pilot-towards-embedded-evaluations