Inside, on agreed terms
The change is real. Apollo's own post notes its previous evaluations of frontier models took place after post-training and shortly before public release. This pilot gives Apollo access during training, evaluation and internal deployment, rather than limiting the engagement to a pre-release model assessment. Apollo's brief is specific: whether a model attempts to subvert its own alignment or control mechanisms, or those of its successors.
The terms are equally specific, and Apollo published them itself. That is the strongest feature of this announcement. Rather than asking readers to trust the word "independent," Apollo set out three requirements for a credible embedded evaluation — access, disclosure and meta-transparency — and then stated plainly that this pilot meets the third, and is "an ambitious first step" toward the first two, without yet fulfilling them.
Where the access stops
Start with scope. Apollo's post states that its assessment covers only the claims the two sides agreed, and only the models and time periods agreed for them. Within that scope, OpenAI may still withhold items where it deems withholding reasonably necessary for technical infeasibility, security risks or IP protection. It has agreed not to withhold material, or access to staff, simply because the material is unfavorable, and it must tell Apollo that withheld items exist and give written reasons. Expansion of scope can be requested and refusals escalated, but Apollo notes expansion is not guaranteed.
The settled access is also narrower than the headline suggests. What OpenAI has committed to so far — training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and queries against intermediate checkpoints — covers the first of Apollo's four questions, training incentives. Apollo says the detailed access arrangements for the other three questions, covering evaluations, internal deployment and detectability, remain to be agreed.
Staff access follows the same pattern. Apollo may request staff with specific expertise and suggest individuals, but OpenAI selects the participants, formal conversations are logged, and if an employee outside the authorized group raises confidential information, Apollo must stop the discussion and route any follow-up through its designated OpenAI contact. Evaluators work on OpenAI hardware, on channels that can be monitored, and by default have no access to the general internal channels an OpenAI safety researcher would use. Apollo's own footnote makes the cost of that explicit: informal conversations and chat channels often surface relevant information nobody thought to ask about.
What readers will and will not see
Apollo will publish two reports, with its methods, conclusions and limitations. OpenAI reviews them first and may request redaction of confidential information, including evidence that supports Apollo's findings. It may not redact Apollo's verdicts, and it may not request a redaction merely because a finding is unfavorable. The unredacted reports go to the Safety and Security Committee of the OpenAI Foundation's board.
That leaves a gap readers should keep in mind. A public report can carry a conclusion whose supporting evidence has been removed as sensitive IP, whether or not that evidence contributed to the finding. Apollo can say that a redaction happened and whether it affected the conclusions, and it can describe withheld material, delayed access and any pressure it believed shaped its work, without prior review by OpenAI. Those meta-transparency rights are what Apollo calls the minimum standard, and on its account this pilot meets them.
There is a second filter on anything learned from staff. Claims based on employee discussions, including formal interviews with authorized staff, can be published only with OpenAI's formal attestation, unless Apollo establishes them independently. Apollo can disclose that an attestation was declined and explain the effect on its conclusions, but not the underlying information.
Four groups, four different arrangements
Apollo has published detailed terms, which should not be assumed to govern the other engagements. According to TechCrunch and Unite.AI's accounts of OpenAI's announcement, Goodfire will embed interpretability researchers to analyze model internals during training runs. METR's first assessment looks for misaligned behavior in OpenAI's largest reinforcement-learning run to date; a second, broader monitor assessment is preliminarily agreed but still pending agreement on scope. Transluce starts with the quality and coverage of OpenAI's post-training alignment evaluations, with an intention to run a follow-on assessment of about eight weeks in early 2027. Payment also splits the group, per the same accounts: Apollo and Goodfire accept compensation that is not contingent on results, while METR and Transluce do not accept financial compensation.
OpenAI presents the program as part of its commitments to independent assessment of its safety controls. Nothing in today's announcement is a finding about any model. There are no results, and no date for Apollo's first report has been published in the sources read for this piece.
What to watch
Three useful follow-up questions will show whether this becomes a standard or stays a pilot. First, whether the access arrangements for Apollo's remaining three questions are agreed before the four months run out. Second, how much of the two public reports survives redaction, and whether Apollo uses its meta-transparency rights to say so in specific terms. Third, whether other labs adopt Apollo's published criteria — access, disclosure, meta-transparency — as the test an embedded evaluation has to pass, rather than treating physical presence inside a lab as the test in itself.
Sources
- [1] Apollo Research, "A Pilot Towards Embedded Evaluations", November 16, 2026 (read first-hand by ALFW)Read source
- [2] OpenAI, "OpenAI Embeds Apollo, METR, Transluce and Goodfire Researchers for First Phase of Model Assessments", November 16, 2026 (announcement read via TechCrunch and Unite.AI; listed without a URL because the desk did not confirm the openai.com address)
- [3] TechCrunch, "OpenAI will embed safety researchers to investigate AI subversion", November 16, 2026Read source
- [4] Unite.AI, "OpenAI Signs Four Independent Assessors for Embedded Safety Reviews", November 16, 2026Read source