Most AI safety reports are written by the lab being judged. OpenAI is now trying something different, and it’s worth watching whether it holds up. The company announced a first phase of “embedded assessments,” with four outside groups working inside its systems: Apollo Research, METR, Transluce and Goodfire.
Embedded means access well beyond API testing. Depending on the evaluation, assessors can see model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and the engineers behind the work. That is much closer to an audit than a red-team contract.
What each assessor will do
Apollo Research will test whether models could covertly work against their developers, and whether OpenAI’s safeguards would catch it. It will check four claims: training did not incentivize subversion, strong evaluations and red teaming do not elicit it, models have not attempted it during internal use, and reasoning behind it would be detectable.
METR gets two jobs. The first is assessing misaligned behavior in OpenAI’s largest reinforcement learning run to date, including failures existing tests may have missed. The second is a broader monitor assessment, still pending agreement on scope. Transluce will review the quality and coverage of OpenAI’s post-training alignment evaluations, with a follow-on of about eight weeks planned for early 2027. Goodfire will embed interpretability researchers to analyze model internals, mainly to help OpenAI’s own staff with training run monitoring and debugging. It says it will publish findings too.
The ground rules
Assessors who need sensitive information work on site, on OpenAI devices, under at least the same security standards as employees. Where sharing model weights or architecture is too risky, OpenAI may pair assessors with its own researchers to run analyses instead. That’s a real limit on independence, and the company doesn’t hide it.
Oversight sits with the Safety and Security Committee of the OpenAI Foundation’s board, which receives unredacted reports. Disagreements go to the Safety Advisory Group and executives. So the escalation path ends inside the company.
Does it matter?
Somewhat, and more than most safety announcements. Anthropic and Google DeepMind both work with outside evaluators, and METR and Apollo have tested frontier models from several labs. But testing a finished model is one thing. Looking at training runs and internal monitoring is another, and that’s where this goes further.
The caveats are plain. This is a pilot. Scope on some assessments is unsettled. OpenAI picked the assessors, the scope and the access, and the company says it is tied to commitments under the White House Accord on Super Intelligence. What matters is whether reports come out unedited and whether any of them are uncomfortable. If they are, this becomes a template other labs get pressed to copy. If not, it’s a well-designed press release.




