METR proposes public reviews of frontier AI loss-of-control risks

METR has proposed a framework for public reviews to address potential loss-of-control risks in frontier AI development. The initiative aims to provide the public and policymakers with independent evidence regarding whether developers can effectively manage their internal AI systems as they become more capable.

The proposal focuses on three primary areas: alignment, control, and the acceleration of AI research and development. Alignment reviews would investigate internal incidents and whether training processes might conceal misalignment. Control assessments would test the ability of internal AI agents to cause undetected harm, while the third track would measure how much AI automation contributes to rapid capability growth.

METR emphasizes that these reviews are not intended to provide safety guarantees or establish formal oversight. Instead, the goal is to increase transparency by sharing evidence gathered through assessments. Previous work by METR includes an agreement with Anthropic to investigate unauthorized access incidents involving Claude models and a joint investigation with Redwood Research regarding an incident involving OpenAI and Hugging Face.