OpenAI has signed agreements with Apollo Research, Goodfire, METR and Transluce for a first phase of embedded assessments that can include access to unreleased models, checkpoints, source code, training data and internal monitoring systems. The program is designed to give outside specialists more evidence than a conventional public benchmark or short red-team exercise.
The word “independent” still needs qualification. OpenAI and each assessor will agree on scope and security arrangements. Sensitive work will take place on company devices at OpenAI offices, and OpenAI says it can redact confidential information or request a delay when publication could create a security or safety risk.
The OpenAI embedded assessments are therefore neither an internal audit nor unrestricted outside access. Their credibility will depend on what each team can inspect, which findings reach the public and how disagreements are disclosed.
What OpenAI embedded assessments can access
OpenAI says assessors may receive some of the resources used by its own researchers, depending on the evaluation. The list includes model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation and discussions with the people responsible for the work.
That depth matters because deployed-model testing sees only behavior available through a finished interface. Checkpoints can show how behavior changes during training. Monitoring records can show what internal systems flagged. Training interventions and evaluation rollouts can reveal whether a safeguard was tested only at the end or shaped the model earlier.
Access will not be identical for every group. The announcement describes different questions and evidence for each engagement rather than one common certification test.
Apollo will test claims about model subversion
Apollo Research is expected to examine whether models could work covertly against their developers’ intentions and whether safeguards would detect that behavior. OpenAI lists four claims for the assessment: training did not incentivize subversion; red teaming did not elicit it; models did not attempt it during internal use; and reasoning associated with subversion would be detectable.
For the first stage, Apollo is slated to receive access to selected public and unreleased models, checkpoints, relevant training data, evaluation results, intervention records, monitoring systems and flagged behavior.
The wording matters. An assessment can test specified claims within selected systems and conditions; it cannot prove that no future model, tool configuration or deployment will ever show deceptive behavior.
Goodfire will examine internal model mechanisms
Goodfire plans to embed interpretability researchers and use its platform to analyze model internals. The stated objective is to improve monitoring and debugging during training and support alignment work.
OpenAI says the primary audience for this assessment is its own researchers, although Goodfire will publish lessons and findings as the collaboration proceeds. That creates a distinction between internal utility and public accountability: a project can improve engineering even if only part of the evidence is released.
METR will assess alignment and monitoring
METR is expected to examine misaligned behavior in OpenAI’s largest reinforcement-learning run to date, including failures that existing tests or monitors may have missed. Access is expected to include selected checkpoints, training and evaluation rollouts, and monitoring records.
OpenAI also describes a possible broader follow-up to an earlier monitor stress test. The prior exercise, documented by METR, is useful context because it shows the difference between evaluating the model and evaluating the system intended to watch the model.
The announcement notes that the first alignment scope is narrower than METR’s own broader proposal. That is an important limitation to preserve when the eventual report is published.
Transluce will examine evaluation quality and coverage
Transluce will initially assess OpenAI’s post-training alignment evaluations: what they test, how well they cover relevant behavior and whether the resulting evidence supports the safety claims built on top of them.
The planned access includes selected public and unreleased models, evaluation results, behavior data, safety-case documentation and researcher discussions. OpenAI and Transluce also say they intend a broader eight-week follow-up in early 2027, but that future work has not yet produced findings.
This “evaluation of the evaluations” is important because a green result is only as informative as the test behind it. Future & Fact’s guide to AI benchmarks explains why task selection, prompts, tools, scoring and omitted failure modes shape what a number can establish.
How independent is an embedded assessor?
Independence is not a binary label. At least six dimensions should be visible in the final reports:
- Scope: Did the assessor choose the questions, or test claims selected by OpenAI?
- Access: Which models, checkpoints, records and employees were available, and what was withheld?
- Method: Could the assessor change its approach when new evidence appeared?
- Funding: Who paid, and was compensation linked to a particular conclusion?
- Publication: Who could redact, delay or prevent details from being released?
- Disputes: Can readers see unresolved disagreements between the lab and assessor?
OpenAI says Apollo and Goodfire will be compensated, while METR and Transluce are not being paid for the described work, though they may receive resources such as compute credits. The company says payment is not contingent on conclusions.
That disclosure helps readers evaluate conflicts, but it does not answer every question. The final assessment should explain material support, staff access and restrictions in enough detail for others to judge the process.
Redactions and responsible disclosure can be legitimate—and limiting
Publishing model weights, exploit details or confidential training information could create real security and intellectual-property risks. A delay can give a company time to fix a vulnerability before outsiders can misuse it.
The same mechanism can reduce transparency if readers see conclusions without the evidence needed to evaluate them. OpenAI says its Safety and Security Committee will receive unredacted reports and can communicate privately with assessors. Disputes can be escalated through internal groups and ultimately to that committee.
The public reports should therefore state what was redacted, why, who approved the restriction and whether the assessor believes the published evidence still supports its conclusion. A responsible-disclosure process is strongest when its limits are themselves documented.
What would count as a useful result
The announcement is a governance commitment, not evidence that a safeguard has passed. The useful outputs will be assessor-authored methods, findings, limitations and the lab’s response.
Readers should look for concrete outcomes: a safety claim supported or rejected, a monitor failure reproduced, an evaluation gap identified, a training intervention changed, or an unresolved disagreement recorded. Vague statements that the collaboration was “valuable” would offer less accountability than a narrow negative result with a clear method.
OpenAI’s approach could establish a stronger model for access to systems that cannot be evaluated fully from the outside. Whether it does so will be determined by the reports—not by the depth of access promised before the work is complete.


