OpenAI has signed agreements with four independent assessors: Apollo Research, METR, Transluce and Goodfire. The company announced it on November 16. As “embedded assessors,” the four get deep access to internal models and systems, depending on the assessment: model checkpoints, source code, training and evaluation data, monitoring systems and time with the researchers responsible. The first phase covers alignment and safeguards during training and internal use. OpenAI ties it to its commitments under the White House Accord on Super Intelligence, where it agreed to independent assessments of its safety controls.
Who checks what
- Apollo Research will look at whether models could covertly work against their developers’ intentions and whether OpenAI’s safeguards would catch it. Among the claims it tests: that strong evaluations and red teaming don’t elicit subversion, and that models haven’t attempted it during internal use.
- METR will assess misaligned behavior in OpenAI’s largest reinforcement learning run to date, including failures existing tests and monitors missed. A second, broader assessment of the monitoring systems is preliminarily agreed, with scope still open.
- Transluce reviews OpenAI’s post-training alignment evaluations for quality and coverage. An eight-week follow-on is planned for early 2027, with a view into how results feed release decisions.
- Goodfire embeds interpretability researchers to analyze model internals during training. Its findings go mainly to OpenAI’s own researchers, though Goodfire plans to publish as it goes.
The rules of access
The protocols OpenAI published alongside are the more telling part. Assessors who need confidential data work on site at OpenAI offices, on OpenAI devices. The Safety and Security Committee of the OpenAI Foundation’s board gets unredacted reports and can talk to assessors privately. Disagreements go first to the Safety Advisory Group and OpenAI executives; if they can’t settle them, the committee decides.
Conversations with employees are off the record. They can inform an assessment but can’t be published or attributed. If an assessor wants to use something from those talks, they can request a statement attested by OpenAI. If OpenAI declines, the assessor may say so publicly. Assessors decide what they publish, but OpenAI keeps the right to redact confidential information and to ask for disclosure delays when findings pose security or safety risks.
The leverage sits in the redactions
In September, OpenAI wrote down its principles for third-party assessments. Now there are names, tasks and procedures. That’s real progress, especially with METR getting into the largest training run. How independent it ends up looking depends on two things OpenAI keeps in its own hands: what gets redacted and how long publication can be delayed. The assessors’ right to say when OpenAI refused to attest to something is the counterweight. Whether it holds up, the first report will show.