From the source
OpenAI says it signed agreements with Apollo Research, METR, Transluce and Goodfire for the first phase of embedded assessments, giving outside organizations sustained access to internal models and systems during training and internal use.
- The access can include model checkpoints, source code, training data and monitoring systems, and OpenAI says it is meant to let assessors examine evidence, run independent evaluations and uncover failures the company may not have anticipated.
- Apollo Research will examine whether models could covertly act against their developers' intentions and whether OpenAI's safeguards would detect such behavior, across four propositions covering training incentives, red teaming, internal use and detectability of reasoning.
- METR's first-phase alignment assessment focuses on what OpenAI describes as its largest reinforcement-learning run to date, including failures existing tests and monitors may have missed, with its initial scope confined to that run; a separate METR monitor assessment remains pending mutual agreement on scope.
- Transluce will begin by examining the quality and coverage of OpenAI's post-training alignment evaluations, and the two sides share an intention to conduct an approximately eight-week follow-on assessment in early 2027.
- Goodfire will embed interpretability researchers and use its platform for training-run monitoring, debugging and alignment, with OpenAI researchers as the primary audience for findings.
- Assessors handling sensitive confidential information will work on-site at OpenAI offices using company devices, and oversight runs to the Safety and Security Committee, which receives unredacted reports and can speak privately with assessors.