OpenAI Signs Four Independent Assessors for Embedded Safety Reviews
OpenAI signed Apollo Research, METR, Transluce and Goodfire on November 16, 2026 for embedded safety assessments with access to checkpoints, training data and monitoring systems.
Why it matters
- OpenAI signed Apollo Research, METR, Transluce and Goodfire for its first embedded assessments, announced November 16, 2026.
- Assessors get access that can include model checkpoints, source code, training data and monitoring systems.
- METR's first phase targets OpenAI's largest reinforcement-learning run to date.
- An approximately eight-week follow-on assessment is planned for early 2027.
- The Safety and Security Committee receives unredacted reports and can speak privately with assessors.
OpenAI has signed Apollo Research, METR, Transluce and Goodfire to conduct its first "embedded assessments," giving outside researchers sustained access to model checkpoints, source code, training data and monitoring systems. The company announced the agreements on November 16, 2026. The four organizations will evaluate alignment and safeguards during training and internal use, not just after release — a structural shift in how frontier-lab safety work gets checked.
The stakes are straightforward. OpenAI's own monitors can miss failures, and the company admits it. The stated purpose of sustained access is to let independent assessors examine evidence in context, run their own evaluations, and surface failures OpenAI didn't anticipate. Public reporting from these assessments, OpenAI says, should give people a clearer view of how laboratories evaluate and manage risks.
What does each assessor actually do?
Each organization has a distinct mandate, and the differences matter.
Apollo Research will probe whether models could covertly act against their developers' intentions and whether OpenAI's safeguards would catch such behavior. Its two-stage assessment covers 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 such behavior would be detectable
For the first stage, OpenAI will hand over selected public and unreleased models, training checkpoints, relevant training data, evaluation results, training-intervention records, monitoring systems and flagged behaviors. Apollo also gets access to OpenAI researchers for context, plus the tools and compute to run independent tests.
Goodfire takes a different angle. It will embed researchers specializing in interpretability — the analysis of model internals — and apply its platform to training-run monitoring, debugging and alignment. OpenAI's own researchers are the primary audience for Goodfire's findings, which they'll use to make safety improvements. Goodfire will also publish lessons as the collaboration develops.
METR has the narrowest and arguably most technical scope. Its first-phase alignment assessment targets what OpenAI describes as its largest reinforcement-learning run to date, including failures that existing tests and monitors may have missed. Within the selected run and the agreed access, resources and safety arrangements, METR would independently set its own priorities and iterate on searches, analyses and elicitation experiments. Its access includes relevant checkpoints, training and evaluation rollouts, and misalignment-monitoring records. The initial scope stays confined to that single run rather than a broad survey of misalignment.
A separate METR monitor assessment is less settled. The two organizations have preliminarily agreed to pursue a broader exercise, but it remains pending mutual agreement on scope.
Transluce starts by examining the quality and coverage of OpenAI's post-training alignment evaluations. Its access covers selected public and unreleased models, evaluation results, supporting behavioral data, safety-case documentation, researchers, tools and compute.
The four mandates complement each other: covert behavior, model internals, a deep dive into one training run, and the quality of post-training checks.
What comes after the first phase?
The parties also share an intention to run an approximately eight-week follow-on assessment in early 2027. That work would broaden access to additional models, deepen access to alignment evaluations, and provide information about how evaluation results feed into release decisions. OpenAI's announcement frames this as a shared intention — deliberately weaker language than the committed first phase.
How much access is real access?
Every engagement starts with an agreed scope plus the access, resources and safety arrangements the work requires. The protocols give assessors freedom to follow evidence, refine their methods, and propose further assessments when findings warrant it, with a defined process for requesting additional access.
There are limits. Where direct access could expose sensitive intellectual property — model weights or architecture, for instance — OpenAI will consider alternatives, including pairing assessors with its own researchers to run analyses. Assessors handling sensitive confidential information will work on-site at OpenAI offices using company devices. They're expected to follow at least the same security standards as OpenAI employees, with stricter requirements possible for some experiments.
Who oversees the overseers?
Oversight runs up to the Safety and Security Committee, an independent oversight committee of the OpenAI Foundation's Board. The committee receives unredacted reports and can speak with assessors during the work — including privately.
Disputes follow a defined ladder. They go first to the Safety Advisory Group and OpenAI executives, who must resolve them within a specified period. The committee must be notified promptly when a disagreement is escalated, and unresolved matters land with the committee itself.
Employee conversations raise their own transparency problem. Those discussions are off the record: they can inform assessments, but assessors cannot publish them or attribute them to employees. There is a narrow workaround. Assessors can request an OpenAI-attested statement verifying, correcting or clarifying something learned in those conversations. They may then publish the attested statement — or disclose that OpenAI declined to attest. That creates a limited public route for information gathered confidentially.
Can the assessors publish freely?
Mostly. Assessors maintain independence in publishing their assessments, but OpenAI retains the right to redact confidential information. The company may also request responsible-disclosure delays when findings could create security or safety risks, and it can use such delays to remediate unresolved issues. OpenAI says public reports should explain both findings and their limitations, while more detailed confidential reports give oversight bodies the evidence they need to act.
The money differs by organization, and the split is notable:
- Apollo Research and Goodfire accept compensation that is not contingent on assessment results
- METR and Transluce accept no financial compensation, though they may take API credits or compute for use in their assessments
Conflict-disclosure and cool-off processes are still being developed with the assessors. OpenAI asks participating third-party staff to observe a brief cool-off period before embedding with another AI developer, to reduce the risk of inadvertently disclosing confidential information.
OpenAI expects early technical-access, space and security challenges, says the protocols will evolve, and describes longer-term embedded engagements as the objective. Future assessments under consideration would cover safety-framework compliance, organizational culture and processes.
Why this is a governance test, not just a research project
The first phase sets up a concrete experiment: can deep internal access plus independent publication actually produce useful public evidence about failures that internal tests miss? The test cuts both ways. It requires the assessors to find things OpenAI's own teams cannot, and it requires OpenAI to hold to the agreed scopes, on-site security controls, redaction rights and possible disclosure delays without hollowing out the process.
If the model works, it could become a template for how frontier labs open themselves to outside scrutiny during training rather than after deployment. If it stalls — through redactions, delayed disclosures or contested scope — the experiment will have shown where the limits of voluntary embedded oversight lie.
Source: unite.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- Apollo Research Embeds Six Evaluators Inside OpenAI for Subversion Pilot
- OpenAI Lays Out Rules for Independent AI Safety Audits
- OpenAI details how external testers probe its frontier models
- OpenAI and Hugging Face reveal findings from model evaluation security incident
- OpenAI Reveals Its Layered System for Catching Misaligned AI Agents