OpenAI announced on November 16, 2026, that Apollo Research will assess whether its models could covertly work against developers’ intentions—and whether safeguards could detect that behavior. The assessment covers four claims about training, evaluations, models’ internal use, and the detectability of their reasoning. OpenAI’s announcement also describes the resources Apollo is set to receive and how publication and oversight will work.

An assessment’s value depends on what it can examine, how it can report findings, and who can review them independently.

What Apollo Research will assess

OpenAI’s four stated claims address different possible paths to subversion: whether training incentivized it, whether strong evaluations or red-teaming can elicit it, whether models have attempted it during internal use, and whether the reasoning behind it would be detectable. Here, subversion means a model covertly working against its developers’ intentions.

Together, the questions cover both behavior and detection. Apollo’s remit includes examining whether the behavior might emerge under particular training or testing conditions, as well as whether monitoring could identify it. They are assessment questions, not reported findings.

What access and publication rules apply

For Apollo’s first assessment stage, OpenAI says access is set to include selected public and unreleased models and training checkpoints; relevant training data and evaluation results; records of training interventions; monitoring systems and flagged behaviors; researcher context; and tools and compute for independent testing. OpenAI says each engagement has an agreed scope, resources, and safety arrangements, with room for assessors to follow evidence and refine their methods within that scope.

OpenAI says assessors retain the ability to publish their assessments. It may redact confidential information and request delays for responsible disclosure when findings could pose safety or security risks. Discussions with OpenAI employees are off the record and cannot be published or attributed to individuals. Assessors may request an attested statement through a separate process and disclose if OpenAI declines to attest it.

The OpenAI Safety and Security Committee receives unredacted assessment reports and may communicate privately with assessors. OpenAI’s stated process sends disagreements to its Safety Advisory Group and executives; unresolved disputes go to the committee. That gives the committee access to the full reports even when public disclosure is limited.

How the four assessors’ assignments differ

OpenAI assigned separate work to Apollo Research, METR, Transluce, and Goodfire. Their remits and compensation arrangements are not the same.

AssessorAnnounced focusCompensation disclosure
Apollo ResearchModel subversion and whether safeguards can detect it, across the four claims about training, evaluations, internal use, and detectability.OpenAI says Apollo is paid, with compensation not contingent on results.
METRAlignment assessment of OpenAI’s largest reinforcement-learning run to date, including possible failures missed by existing tests and monitors.METR does not accept financial compensation for the assessment, though it may accept compute or API credits.
TransluceQuality and coverage of post-training alignment evaluations. OpenAI also expressed intent to arrange a follow-on assessment of about eight weeks in early 2027.Transluce does not accept financial compensation for the assessment, though it may accept compute or API credits.
GoodfireInterpretability work for training-run monitoring, debugging, and alignment. OpenAI says the main audience is its internal researchers, while Goodfire will also publish findings.OpenAI says Goodfire is paid, with compensation not contingent on results.

The distinction matters: Apollo’s work focuses on subversion, while the other three assignments address alignment evaluations, a reinforcement-learning run, and interpretability. The compensation terms also distinguish a fee from possible compute or API credits.

OpenAI cites breach of trust in response to former safety researchers’ letter