OpenAI Signs Agreements With Four Independent Firms for Embedded Safety Assessments
Apollo Research, METR, Transluce and Goodfire will get scoped access to internal models and systems to assess alignment and safeguards in a pilot program.

OpenAI announced Monday that it signed agreements with four independent evaluators to assess the behavior of its artificial-intelligence models and the safeguards around them, giving outside researchers access to some internal models and systems.
Apollo Research, METR, Transluce and Goodfire will take part in the first phase of what OpenAI calls embedded assessments. The work will examine model alignment and safeguards during training and internal use. The announcement is the start of an assessment program, not an independent finding that the models are safe.
Access will vary with each project. OpenAI says it can include model checkpoints, source code, training and evaluation data, monitoring systems, technical documents and discussions with researchers and engineers.
Apollo Research will examine whether models could covertly work against their developers' intentions and whether safeguards would detect that behavior. Goodfire will use interpretability methods to study model internals, with an emphasis on training-run monitoring, debugging and alignment.
METR plans to assess misaligned behavior in OpenAI's largest reinforcement-learning run to date, including problems the company's existing tests or monitors may have missed. The two organizations also intend to conduct another assessment of monitoring systems, though that work remains subject to agreement on scope.
Transluce will review the quality and coverage of OpenAI's post-training alignment evaluations. The organizations also intend to conduct a broader follow-on assessment of about eight weeks in early 2027.
OpenAI says each engagement begins with an agreed scope and arrangements for access, resources and safety. The company acknowledges that technical access problems, confidentiality rules and security protocols can create limits and says the processes will evolve as the work continues.
The company says the assessors will share findings under project-specific agreements and that the program is intended to give the public a clearer view of how risks are evaluated. The different projects are not a single certification process, and Monday's announcement did not present completed assessment results.