Dev News Daily ENDE

OpenAI signs Apollo Research, METR, Transluce and Goodfire for embedded safety assessments

OpenAI said on 16 November that it has signed agreements with four independent assessors - Apollo Research, METR, Transluce and Goodfire - for a first phase of what it calls embedded assessments. Assessors receive, depending on what each evaluation needs, some of the same internal resources its researchers use: model checkpoints, source code, training and evaluation data, monitoring systems, documentation and discussions with the people responsible. The first assessments focus on alignment and safeguards for models during training and internal use.

Apollo Research will test four claims in two stages: that training did not incentivise 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. METR will assess misaligned behaviour in OpenAI's largest reinforcement learning run to date, including failures existing tests and monitors may have missed; a second, broader monitor assessment is agreed only preliminarily, pending scope. Transluce will examine the quality and coverage of OpenAI's post-training alignment evaluations, with a follow-on of about eight weeks planned for early 2027. Goodfire will embed interpretability researchers for training-run monitoring and debugging; its main audience is OpenAI's own researchers, though it will publish learnings.

The post sets out the protocols. Assessors who need sensitive information work on site at OpenAI's offices on OpenAI devices. The Safety and Security Committee of the OpenAI Foundation's board receives unredacted reports and can speak with assessors privately; disagreements go to the Safety Advisory Group and executives and, if unresolved, to the committee. Conversations with employees are off the record, but assessors can request statements that OpenAI attests to, and may publish either the statement or the fact that OpenAI declined to attest. Assessors keep independence in publishing, while OpenAI retains the right to redact confidential information and may ask for disclosure delays to fix findings. Where access could expose weights or architecture, OpenAI may pair assessors with its own researchers instead.

OpenAI links the work to its commitments under the White House Accord on Super Intelligence, where it agreed to independent assessments of its safety controls.

OpenAI signs Apollo Research, METR, Transluce and Goodfire for embedded safety assessments
OpenAI signs Apollo Research, METR, Transluce and Goodfire for embedded safety assessments — Dev News Daily

Why it matters

The value of such an assessment depends on the terms more than on the names. Here they are written down: who sees the unredacted report, what happens to a disagreement, and that a refusal to attest can itself be published. The limits are written down too - redaction rights, delays and the fact that some access is still pending agreement.