← Monday, November 16 deck

03Research

OpenAI Gives Apollo Research, METR, Transluce and Goodfire Deep Access to Internal Models

We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards. These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems.

Context

Why it matters

Independent testers with access to checkpoints, code and monitoring can catch alignment failures that a lab's own evaluations miss, and the findings will test OpenAI's safety claims in public.

OpenAI announced the first phase of its embedded assessments program on November 16, saying it has signed agreements with Apollo Research, METR, Transluce and Goodfire. The company said the work gives outside experts deep access to internal models and systems to investigate alignment and test safeguards. OpenAI published a blog post the same day that describes the scope, the access it will grant, and the rules for publication.

Deep dive

OpenAI said it is working with Apollo Research, METR, Transluce and Goodfire on embedded assessments. In a November 16 blog post, the company said assessors receive, based on what each evaluation needs, some of the same internal resources its researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the staff responsible for the work.

The first assessments focus on alignment and safeguards for models during training and internal use. OpenAI said Apollo Research will assess whether models could covertly work against their developers' intentions and whether safeguards would detect that behavior. METR will assess misaligned behavior in OpenAI's largest reinforcement learning run to date, and a broader monitor assessment is planned pending agreement on scope. Transluce will review the quality and coverage of post training alignment evaluations, with a later assessment of about eight weeks planned for early 2027. Goodfire will embed interpretability researchers to analyze model internals, with findings aimed first at OpenAI's own researchers and later at public writeups.

OpenAI said assessors who need sensitive information will work on site on company devices. The Safety and Security Committee of the OpenAI Foundation's Board will receive unredacted reports and can speak with assessors privately. Assessors keep independence in publishing, but OpenAI retains the right to redact confidential information and may request delays so it can act on findings that pose security or safety risks. The company said this work also supports its commitments under the White House Accord on Super Intelligence. It disclosed that Apollo Research and Goodfire accept payment that is not tied to results, while METR and Transluce do not accept payment but may accept compute.

What to expect next

OpenAI said public reporting from this first phase should give a clearer view of how it identifies and manages risks, and that it expects the work to grow into longer, regular embedded engagements. Watch for Transluce's planned later assessment in early 2027, for whether METR and OpenAI agree on a broader monitor review, and for any public findings from Apollo Research and Goodfire. The company also said it is exploring later assessments of compliance with its safety frameworks and of organizational culture and processes.