OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire for assessments during model development, The Defiant reported on 16 November. METR will examine behaviour at odds with developers’ intentions during what OpenAI describes as its largest reinforcement learning run to date.
OpenAI will give METR model snapshots, transcripts and monitoring records throughout training. Under OpenAI’s plan, METR can choose what to investigate and adjust its experiments within agreed limits on access, resources and safety. The pilot covers one run and excludes misalignment incidents OpenAI has already disclosed.
Checking whether an answer follows instructions could benefit if those experiments find behaviour that OpenAI’s existing tests or monitors missed. The assessment is being conducted during training, rather than after the run is complete.
METR plans to publish findings and an account of its process, although OpenAI can remove confidential information and request publication delays for safety or security reasons. OpenAI’s Safety and Security Committee will receive unredacted reports and can speak privately with assessors. A separate review of OpenAI’s monitoring systems remains tentative, with its scope still to be agreed.