OpenAI has signed agreements with four independent assessors — Apollo Research, METR, Transluce and Goodfire — for a first phase of embedded assessments, a deeper form of outside review in which experts work inside the lab's systems over an extended period.
The assessors will get access to some of the same internal resources OpenAI's own researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems and technical documentation, plus conversations with the researchers and engineers behind the work. The initial focus is alignment and safeguards for models during training and internal use.
Each team has a different brief. Apollo Research will test whether models could work against their developers' intentions without being caught. Goodfire will embed interpretability researchers to monitor training runs and alignment. METR will run two assessments, one on alignment in OpenAI's largest reinforcement learning run to date and one on monitoring. Transluce will examine the quality and coverage of post-training alignment evaluations, with a longer follow-on assessment planned for early 2027.
Oversight sits with OpenAI's Safety and Security Committee, which will receive unredacted reports and can speak privately with assessors. Assessors keep their independence in publishing findings, though OpenAI can redact confidential information and request delays where a finding poses a security risk.
The work supports OpenAI's commitments under the White House Accord on Super Intelligence, and the company says it expects the first phase to grow into longer-term embedded engagements.





