On November 16, 2026, OpenAI announced agreements with Apollo Research, METR, Transluce, and Goodfire for its first phase of embedded assessments. Each organization has a distinct brief, and access is set according to the needs of each assessment, OpenAI said.

To understand what this program can examine, look at each team’s assigned scope, how employee conversations are handled, and who receives the reports.

What will each organization examine?

The four organizations are examining different parts of model alignment and safeguards during training and internal use. The scope ranges from tests of possible model subversion to reviews of alignment evaluations and interpretability work.

OrganizationAnnounced focusScope or timing detail
Apollo ResearchFour claims about model subversionTests whether training incentivized subversion, whether evaluations or red-teaming elicit it, whether models attempted it during internal use, and whether reasoning behind it would be detectable. The work is planned in two stages.
METRMisaligned behavior in a selected reinforcement-learning runThe first alignment assessment focuses on a run OpenAI describes as its largest to date. OpenAI says the scope is narrower than METR’s proposed broad survey of misalignment. A separate, broader monitor assessment is preliminary, pending agreement on scope.
TransluceQuality and coverage of selected post-training alignment evaluationsOpenAI and Transluce intend an approximately eight-week follow-on in early 2027, with broader access.
GoodfireInterpretability, training-run monitoring, debugging, and alignmentOpenAI researchers are the primary audience for the findings; Goodfire will also publish learnings.

How does access work?

OpenAI says each engagement has an agreed scope, access, resources, and safety arrangements. Depending on the assessment, evaluators may receive selected model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with relevant researchers and engineers. The protocol does not describe the same access package for every team.

Evaluators may request additional access as their work develops. OpenAI says it may consider alternatives when direct access could expose sensitive intellectual property. For sensitive work, evaluators work onsite using OpenAI devices and follow at least employee-equivalent security standards.

Employee conversations are authorized by OpenAI and held off the record, so they can inform an assessment but cannot be published or attributed to employees. Evaluators may request OpenAI-attested statements for publication, or disclose that the company declined to attest to a statement.

What can evaluators publish, and who oversees the work?

Evaluators retain the ability to publish findings, while OpenAI reserves the right to redact confidential information and may request delays for responsible disclosure. The company also says Apollo Research and Goodfire accept compensation that is not contingent on assessment results. METR and Transluce accept no financial compensation for these assessments, though they may accept API credits or compute.

The Safety and Security Committee of the OpenAI Foundation’s Board receives unredacted reports and can speak privately with evaluators. Disputes first go to the Safety Advisory Group and OpenAI executives; unresolved matters are referred to the committee.

Why has evaluator access drawn criticism?

Sydney Von Arx, founder of AI-safety research group Nightingale, told NPR that only two evaluators seemed able to speak with employees and that evaluators lacked access to internal Slack. NPR also reported that Apollo’s process requires OpenAI to be involved before substantive conversations with employees who are not pre-approved. These are criticisms reported by NPR and attributed to Von Arx and its reporting—not findings from an assessment. NPR reported the criticism and OpenAI’s response.

OpenAI told NPR that evaluators can request conversations with additional employees and described a process for obtaining company statements about information learned informally. Its published protocol specifies that staff conversations are authorized and off the record. Those rules define how the announced assessments are organized; the four workstreams remain distinct, with the Transluce follow-on intended for early 2027.