METR Proposes Public Reviews of Frontier AI Loss-of-Control Risks

METR is proposing public reviews of whether frontier AI developers can control systems used inside their organizations, aiming to give the public evidence about risks without offering a safety guarantee or formal oversight.

The plan covers three areas: alignment, control and AI research-and-development acceleration. METR is prepared to conduct all three assessments with frontier developers and is encouraging companies to involve other external groups and disclose evidence publicly.

Alignment reviews would examine developers’ internal incidents and whether training could conceal more serious misalignment. The reviews would focus on evidence about how systems behave and whether problems inside an organization might point to deeper risks.

Control assessments would test whether internal AI agents could cause severe harm without being detected. The third track would measure how much AI automation contributes to developers’ growth in capabilities.

The proposal is intended to inform the public about the evidence gathered through these reviews. It does not promise that a developer’s systems are safe and does not establish formal oversight. Involving additional external groups and making evidence public are part of the approach METR is urging developers to take.

The proposal follows METR’s work with AI developers. Anthropic signed an agreement with METR to investigate reported incidents involving Claude models’ unauthorized access to third-party systems. METR and Redwood Research also conducted an independent investigation related to OpenAI’s Hugging Face incident.

TokenPost has previously examined AI evaluators’ role as California expanded oversight rules.