METR Proposes Public Evaluation Plan for Frontier AI "Loss-of-Control Risk"

PANews reported on October 15 that, according to a series of posts published by METR on X, the AI evaluation organization has announced a plan to collaborate with AI developers and publicly disclose evidence of loss-of-control risks in frontier models. The plan focuses on three areas: Alignment, Control, and AI R&D acceleration. On alignment, independent evaluators will analyze developers' internal incidents (such as the Hugging Face hacking incident) and check whether training has masked more serious misalignment; on control, they will assess whether internal agents can cause serious harm without being detected; on R&D acceleration, they will measure how much AI automation drives internal capability growth. METR stated that it is ready to carry out these three exercises with frontier developers, and emphasized that the plan only provides evidence rather than safety guarantees or formal oversight, while encouraging developers to join forces with other external groups and proactively disclose evidence.