METR proposes public reports on whether AI developers could lose control
This is a proposal, not a completed assessment. METR does not say which developers, if any, have agreed to these reports; it mentions other ongoing exercises, including an alignment assessment with Anthropic. Access would be at each company's discretion, companies could redact information or stop the engagement, and METR says the exercises are not audits and would not give confident assurance of safety.
What was claimed?
METR proposes that frontier developers let it assess three questions and publish reports on them: how severe the misaligned actions of internal agents could be, whether company monitoring would catch and stop harmful actions, and whether AI-driven research could speed capability growth beyond what safeguards keep up with. Reports would draw on private staff interviews and internal data, and METR says it would aim to publish preliminary reports within weeks.
What does the evidence support?
Documented availability, not performance proof
Most outside evaluation so far has covered models before public release. This proposal targets AI systems used inside developers, where METR's earlier pilot report found that Anthropic's monitoring at the time could be circumvented in several ways.
What are the limits?
This is a proposal, not a completed assessment. METR does not say which developers, if any, have agreed to these reports; it mentions other ongoing exercises, including an alignment assessment with Anthropic. Access would be at each company's discretion, companies could redact information or stop the engagement, and METR says the exercises are not audits and would not give confident assurance of safety.
Claim or evidence date: 2026-10-14. Last verified: .
METR · A proposal for better informing the public about loss-of-control risks ↗
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →