Anthropic Opens Internal AI Safety Systems to METR

Anthropic is giving six researchers from independent evaluator METR employee-level access to assess its models and internal safety processes.

What happened

The METR researchers will receive company laptops, office access and internal tools comparable to those used by Anthropic's own risk teams. The arrangement is designed to let an outside organisation study not only model behaviour but also the processes used to evaluate and manage frontier-model risk. METR will publish its own conclusions, and Anthropic cannot veto publication, although narrowly defined confidential information can be redacted. The initiative goes further than conventional third-party benchmarking, where evaluators often receive limited model access and little visibility into internal systems.

Why it matters

Independent AI safety testing is difficult when developers control access to models, infrastructure and information. Employee-level access gives evaluators a stronger chance of identifying risks that external black-box testing could miss. The publication arrangement is also important because independence depends on evaluators being able to report uncomfortable findings.

The bigger picture

If the model proves workable, embedded independent evaluation could become a template for frontier AI oversight. The central question is whether deep access can be combined with genuine independence and meaningful public disclosure without exposing sensitive security or proprietary information.