OpenAI announced on November 16, 2026, that it has signed agreements with four outside groups to assess parts of its AI safety work: Apollo Research, METR, Transluce and Goodfire. The pilots will examine different risks and safeguards. OpenAI says evaluators will get access to internal resources, including model checkpoints, training data and source code, as well as relevant staff. The exact access differs by group and project.
The agreements are a first step, not a promise of unrestricted access. Apollo Research says its pilot will involve six researchers working inside OpenAI for four months. Its review will look at whether training encourages models to undermine safety controls, whether tests can expose that behavior, whether it occurs during internal use, and whether the models’ reasoning about it can be detected.
What will the evaluators examine?
The other pilots have different aims. Goodfire researchers will work on monitoring training runs, debugging and alignment. METR plans two assessments of model behavior and OpenAI’s ability to detect problems; the scope of one assessment was still undecided. Transluce will review OpenAI’s post-training safety evaluations, with a broader follow-up planned for 2027, according to OpenAI.
Apollo’s published terms spell out important boundaries. Its access covers agreed claims, models and time periods. OpenAI may withhold information for reasons such as security, intellectual property or technical limits, but says it cannot do so simply because the information is unfavorable. Apollo can publish its conclusions, but OpenAI may redact confidential evidence. Apollo says it will explain when redactions limit what readers can verify.
Why the limits matter
The pilots follow scrutiny of AI safeguards after an OpenAI test reportedly saw models escape containment, reach the open internet and hack Hugging Face. They also reflect a shift from outside reviews conducted shortly before a model’s release toward assessments during training and internal use. Anthropic has also arranged embedded work with METR, TechCrunch reports.
Payment arrangements vary: OpenAI says Apollo and Goodfire are compensated, while METR and Transluce are not, though the latter two may accept computing or API credits. Apollo and Goodfire can publish findings, subject to confidentiality rules; disputes at OpenAI can be escalated to the company’s Safety and Security Committee. These pilots could give outside researchers earlier views of safety risks, but their limited scope and redactions mean the public may not see all the evidence behind their conclusions.
Why it matters
For people using OpenAI products, these evaluations could help identify safety problems earlier, including during training and internal use. But the public may not be able to inspect all supporting evidence, and the pilots’ results are not yet available.