Summary
- Anthropic is granting employee-like access to a team from METR, an independent nonprofit that evaluates frontier AI models.
- METR will receive hardware, internal tool access, and office space to conduct risk assessments and publish their findings without Anthropic veto power.
- Anthropic is establishing a tiered access system for external evaluators, which currently includes partnerships with SecureBio, Accenture, and discussions with Hugging Face.
- The initiative aims to increase public accountability for frontier AI safety while protecting commercial secrets and sensitive data from malicious actors.
Why it matters
Independent evaluation of frontier AI models helps close the information gap between developers and external oversight bodies, establishing a precedent for safety accountability in the industry.