OpenAI on November 16, 2026 announced that it has signed agreements with four independent assessors, Apollo Research, METR, Transluce and Goodfire, for what it calls its first phase of embedded assessments. The company said the work is meant to give independent experts deep access to internal models and systems over an extended period, so they can assess model alignment and the effectiveness of safeguards, especially where those questions matter to OpenAI's safety cases.
According to the post, assessors will, depending on what each evaluation needs, receive access to some of the same internal resources OpenAI researchers use. The list includes model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work. OpenAI said the initial assessments focus on alignment and safeguards for models during training and internal use, and that they support commitments under the White House Accord on Super Intelligence to allow independent assessments of its safety controls.
The company described a distinct brief for each organization. Apollo Research is to assess whether models could covertly work against their developers' intentions and whether OpenAI's safeguards would detect that behavior, testing four claims: that training did not incentivize subversion, that strong evaluations and red teaming do not elicit it, that models have not attempted it during internal use, and that reasoning behind subversion would be detectable. For the first stage, OpenAI said Apollo will receive access to selected public and unreleased models and training checkpoints, relevant training data and evaluation results, training intervention records, monitoring systems and flagged behaviors, and researchers who can provide context, along with tools and compute for independent testing.
Goodfire, OpenAI said, will embed interpretability researchers and use the Goodfire platform to analyze model internals, with a focus on training-run monitoring, debugging and alignment. The primary audience for those assessments is internal OpenAI researchers; Goodfire will also publish learnings as the collaboration proceeds. METR is slated for two assessments. In the first, it would assess misaligned behavior in what OpenAI called its largest reinforcement learning run to date, including failures that existing tests and monitors may have missed, with access to sampling relevant checkpoints, training and evaluation rollouts, and misalignment monitoring records for that run. OpenAI said this first-phase scope is less expansive than METR's published alignment-assessment proposal. A broader monitor assessment is described as preliminarily agreed and still pending mutual agreement on scope, after a recent three-week monitor red-teaming exercise with METR.
Transluce is to conduct an initial assessment of OpenAI's post-training alignment evaluations, looking at their quality and coverage, with access to selected public and unreleased models, the evaluations and relevant results, supporting behavior data and safety-case documentation, and researchers who can provide context. OpenAI and Transluce also share an intent, the post said, to conduct a follow-on assessment of about eight weeks in early 2027, with broader access, including information on how evaluation results inform model release decisions.
OpenAI said it has developed protocols with input from the assessors. Assessors who need access to sensitive confidential information will work on site at OpenAI offices and be provided OpenAI devices. The Safety and Security Committee of the OpenAI Foundation's board, described as an independent oversight committee, will have visibility into the assessments, receive unredacted reports, and be able to speak privately with assessors; unresolved disagreements can be referred to that committee. Assessors keep independence in publishing, OpenAI said, but the company will retain the right to redact confidential information and may request responsible-disclosure delays. Employee discussions are off the record, with a separate attestation pathway for statements OpenAI is willing to confirm. In a footnote, OpenAI said Apollo Research and Goodfire accept financial compensation that is not contingent on assessment results, while METR and Transluce do not accept financial compensation but may accept API credits or compute for use within the assessments.
The announcement is OpenAI's description of agreements and intended access. The assessors' own findings have not been published as part of this first-phase notice.