OpenAI Pilots Embedded Outside Safety Reviews to Test Frontier Model Risks
OpenAI has agreed to let outside groups Apollo Research, Goodfire, METR and Transluce assess its models and development processes from within. Apollo will place six researchers inside the company for four months to study whether training incentives could lead models to undermine alignment or control. Goodfire plans to apply interpretability tools during training runs, while METR will probe dangerous behaviors that OpenAI's internal tests might miss. Transluce's six-week review will cover selected models, internal evaluations and raw chain-of-thought outputs, but without employee-level access. Apollo has cautioned that the arrangement does not yet count as a full embedded evaluation, since access and disclosure remain restricted.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in