Four outside groups will get a look inside OpenAI’s training and safety machinery. OpenAI announced on 16 November that it signed agreements with Apollo Research, METR, Transluce and Goodfire for what it calls embedded assessments. The company says published findings from such work give the public a clearer picture of how labs test and manage risk as their systems grow more capable.

The access is the novel part. Depending on the assessment, outsiders can use resources that OpenAI’s own researchers rely on: checkpoints, source code, training data, evaluation data, monitoring systems, technical documents, and conversations with the engineers responsible. The first round targets alignment and safeguards while models are being trained or used inside the company. OpenAI also ties the effort to the White House Accord on Super Intelligence, under which it committed to have outsiders examine its safety controls.

Each assessor has a different job. Apollo Research will test whether a model could quietly work against its developers, and whether safeguards would catch that. It will examine four claims across two stages. Training never rewarded subversion. Careful testing and adversarial probing fail to surface it. Models have not tried it during internal use. Any reasoning behind it would be visible. Goodfire will embed interpretability researchers to study model internals during training runs, mostly for OpenAI’s own staff, and says it will publish what it learns.

METR has two assessments. The first examines misaligned behavior in the biggest reinforcement learning run OpenAI has done, including failures that current tests and monitors might have overlooked, with METR choosing its own priorities inside the agreed access. A second, broader review of monitors is preliminarily agreed, but its scope is still open. Transluce will judge the quality and coverage of OpenAI’s alignment evaluations after training. Transluce and OpenAI both want a follow-on lasting about eight weeks, early in 2027, with wider access that would reach into how evaluation results shape release decisions.

Now the limits, which OpenAI’s own announcement lays out. Assessors keep independence over what they publish, but OpenAI reserves the power to redact confidential information and can ask for delays on responsible-disclosure grounds. A footnote concedes that METR’s alignment assignment is narrower than the one METR itself proposed, which called for a broad survey of misalignment. Where access could expose the most sensitive material, such as weights or architecture, OpenAI says it may substitute other arrangements, like pairing assessors with its researchers.

Money is the other caveat. Apollo Research and Goodfire accept financial compensation, which OpenAI says does not depend on results. METR and Transluce take no fees, though API credits or compute may be accepted. Nothing in the announcement suggests the pay shapes findings, yet it is a tie a reader should weigh when a report comes back clean.

The process has guardrails. Assessors handling sensitive information work on site, on OpenAI devices, held to security rules no looser than those for employees. The Safety and Security Committee, an independent oversight body of the OpenAI Foundation’s board, gets reports without redactions and can meet assessors privately. Disagreements climb from the Safety Advisory Group and executives to the committee. Conversations with employees are not attributed, and assessors can ask OpenAI to vouch for statements, or say publicly that it declined. Assessor staff observe a short cool-off before joining another AI developer.

The assessments build on a pilot review of OpenAI’s internal monitors, an investigation of the Hugging Face incident, and a recent three-week monitor exercise with METR in which OpenAI opened its codebase and monitors. OpenAI expects longer engagements and is exploring future reviews of how it complies with its safety frameworks, and of its culture and internal processes.

All of this rests on OpenAI’s own account. The announcement says what outsiders may see, not what they have found, and no results are public yet. The first real test comes when an assessor publishes, and whether redaction notes appear beside the findings.

Based on OpenAI’s Safety post “Our first phase of embedded assessments,” published 16 November 2026 at openai.com.