OpenAI Embeds Four Independent Assessors in Safety Reviews
OpenAI signed Apollo Research, METR, Transluce and Goodfire on November 16, 2026, giving them sustained access to assess model alignment, safeguards and failures.
By Amara Osei
5 min read

What's News
- OpenAI announced on November 16, 2026 that it signed first-phase embedded assessment agreements with four organizations.
- Apollo Research’s mandate spans two stages and four propositions on model subversion.
- METR’s initial scope covers one reinforcement-learning run, OpenAI’s largest to date.
- OpenAI and Transluce intend an approximately eight-week follow-on assessment in early 2027.
- The Safety and Security Committee receives unredacted reports and can speak privately with assessors.
By Amara Osei
OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire on November 16, 2026, beginning four embedded assessments of model alignment and safeguards during training and internal use. OpenAI’s agreements give assessors access that can include model checkpoints, source code, training data and monitoring systems.
OpenAI said the access can extend for long periods, particularly around its safety cases. It said sustained access lets assessors examine evidence and context, conduct independent evaluations and uncover failures it may not have anticipated. OpenAI presents public reporting as a way to show how laboratories evaluate and manage model risks.
What will the four assessors examine?
OpenAI tasked Apollo Research with testing whether models could covertly act against developers’ intentions and whether OpenAI’s safeguards would detect such behavior. OpenAI defined Apollo’s two-stage assessment around four propositions:
- Training did not incentivize subversion.
- Strong evaluations and red teaming do not elicit it.
- Models have not attempted it during internal use.
- The reasoning behind it would be detectable.
For stage one, OpenAI will provide selected public and unreleased models, training checkpoints, relevant training data, evaluation results, training-intervention records, monitoring systems and flagged behaviors. Apollo will also receive access to researchers who can provide context, plus tools and compute for independent testing.
Under its OpenAI agreement, Goodfire will embed researchers specializing in interpretability and analysis of model internals. Goodfire will use its platform for training-run monitoring, debugging and alignment. OpenAI researchers are the primary audience for the findings, and they will use them to make safety improvements. Goodfire will also publish lessons and findings as the collaboration develops.
OpenAI said METR’s first-phase alignment assessment would focus on its “largest reinforcement-learning run to date,” including failures that existing tests and monitors may have missed. METR can set priorities independently and iterate on searches, analyses and elicitation experiments within the selected run and agreed access, resources and safety arrangements.
METR will receive relevant checkpoints, training and evaluation rollouts, and misalignment-monitoring records. The initial scope covers only that run, not a broad survey of misalignment. A separate, broader monitor assessment remains preliminary and requires mutual agreement on scope.
Under its agreement with OpenAI, Transluce will begin by examining the quality and coverage of OpenAI’s post-training alignment evaluations. Transluce’s access includes selected public and unreleased models, evaluation results, supporting behavioral data, safety-case documentation and researchers, as well as tools and compute.
What follows the first phase?
OpenAI and Transluce intend to conduct an approximately eight-week follow-on assessment in early 2027. That work would broaden access to additional models, deepen access to alignment evaluations and show how evaluation results inform release decisions. OpenAI described it as a shared intention, distinct from the initial assessment.
How will access and security work?\n
OpenAI starts each engagement with an agreed scope and the access, resources and safety arrangements needed for the work. OpenAI’s protocols give assessors “freedom to follow evidence,” allow them to refine methods and let them propose further assessments when findings warrant them. The protocols also establish a process for addressing additional access needs.
When direct access could expose sensitive intellectual property such as model weights or architecture, OpenAI will consider pairing assessors with its researchers to run analyses. Assessors handling sensitive confidential information will work on-site at OpenAI offices using company devices. OpenAI expects them to follow at least the same security standards as OpenAI employees, with stricter requirements possible for some experiments.
OpenAI expects early technical-access, space and security challenges and says the protocols will evolve. Longer-term embedded engagements are the objective, while future assessments remain under consideration for safety-framework compliance, organizational culture and processes.
Who controls the assessments?
Under OpenAI’s protocols, oversight sits with the Safety and Security Committee, an independent oversight committee of the OpenAI Foundation’s Board. The committee will receive unredacted reports and can speak privately with assessors during the work, while disputes will first go to the Safety Advisory Group and OpenAI executives for resolution within a specified period.
The committee must receive prompt notice when disagreements are escalated. Unresolved matters will go to the committee.
OpenAI keeps employee conversations off the record. They can inform assessments, but assessors cannot publish them or attribute them to employees. Assessors may request an OpenAI-attested statement verifying, correcting or clarifying what they learned. They can publish that statement or disclose OpenAI’s refusal to attest, creating a limited public route for information gathered confidentially.
What can the assessors publish, and do they get paid?\n
Assessors maintain independence in publishing their assessments, while OpenAI retains the right to redact confidential information. OpenAI may also request responsible-disclosure delays to address findings that could create security or safety risks and remediate unresolved issues. OpenAI’s publication rules call for public reports to explain findings and their limitations, while more detailed confidential reports provide oversight bodies with evidence needed to act.
Apollo Research and Goodfire accept compensation that is not contingent on their assessment results, according to OpenAI. METR and Transluce accept no financial compensation, although they may accept API credits or compute for their assessments.
OpenAI and the assessors are still developing conflict-disclosure and cool-off processes. OpenAI asks participating third-party staff to observe a brief cool-off period before embedding with another AI developer, reducing the risk of inadvertently disclosing confidential information.
OpenAI’s first-phase test is whether deep internal access and independent publication can produce useful public evidence about failures missed by internal tests while OpenAI maintains agreed scopes, on-site security controls, redaction rights and possible disclosure delays.
Source: unite.ai
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,136 articles