METR proposes embedded assessments of unreleased AI models’ loss-of-control risks

A post describes METR’s proposal to work with frontier AI developers to make evidence about unreleased models’ loss-of-control risks public. It calls for rapid publication and transparency about what is assessed, redacted and limited. The author says developers could decline access or withhold sensitive information if METR can describe how that affects its findings. The voluntary proposal would not provide “assurance of safety,” the author adds.