OpenAI brings Apollo, METR, Transluce and Goodfire in for embedded safety assessments
Six Apollo staff will work inside OpenAI for four months; Apollo says the terms do not yet meet its full access and disclosure criteria, and OpenAI may redact evidence.
- Safety & alignment
- Labs & people
- Notable
OpenAI said on 16 November that it had signed agreements with four independent assessors, Apollo Research, METR, Transluce and Goodfire, for a first phase of “embedded assessments”. Assessors receive access to some of the internal resources OpenAI’s own researchers use, including model checkpoints, source code, training and evaluation data and monitoring systems, to examine alignment and safeguards during training and internal use. OpenAI tied the work to its commitment, under the White House accord, to independent assessment of its safety controls.
The four assignments differ. Apollo will test four claims in two stages: that training did not incentivise subversion, that evaluations and red-teaming do not elicit it, that models have not attempted it during internal use, and that reasoning behind it would be detectable. METR will assess misaligned behaviour in what OpenAI calls its largest reinforcement-learning run to date, and has preliminarily agreed a broader monitor assessment pending scope. Transluce will review 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, with its main audience inside OpenAI. The Safety and Security Committee of the OpenAI Foundation’s board is to receive unredacted reports, and disagreements are to be escalated to OpenAI’s Safety Advisory Group and executives.
Apollo’s own account is more cautious than OpenAI’s. It says six of its staff will work within OpenAI for four months, on OpenAI hardware, and that OpenAI has agreed not to withhold information or access to staff, or request redactions, simply because findings are unfavourable. But it also says OpenAI “may withhold some in-scope information” on grounds of infeasibility, security or intellectual property, may redact supporting evidence from the two public reports, and chooses which staff Apollo may speak to. Claims drawn from employee conversations need OpenAI’s formal attestation to be published. Apollo said the pilot meets its bar for “meta-transparency”, the right to explain publicly what limited its work, but that it “doesn’t yet fulfill” its criteria for full access and disclosure.
The arrangement follows METR’s earlier work stress-testing OpenAI’s monitoring and its investigation of the Hugging Face incident, both cited by OpenAI. Anthropic had given METR similar access a month earlier. No findings have been published; Apollo said it would report more as the pilot progresses.
In the commentary
What people were saying around this time — external links, from the record's commentary rail.
- 18 November 2026 · Tom Davidson · ForethoughtUpdates from METR report: timelines vs intelligence explosion
- 18 November 2026 · Zvi Mowshowitz · Don't Worry About the VaseQuickly, There’s No Time
- 9 November 2026 · Alexa Pan · Redwood ResearchAstra covertly underperforming on safety research is concerning, even though it’s unlikely to be motivated by ambitious goals