Anthropic Gives METR Employee Level Access to Audit Claude AI Risks

Summary

Anthropic on Thursday embedded six METR employees in its offices, providing company hardware, office space and internal tools and systems with access roughly equal to its own risk assessment teams. Employees are encouraged to treat them as risk colleagues. The move begins implementing chief executive Dario Amodei’s September proposal for resident external auditors, backed by OpenAI’s Sam Altman and SpaceX’s Elon Musk, although OpenAI has not announced its own plan. Anthropic compares the arrangement to resident bank examiners and says increasingly powerful models require stronger safeguards against economic, cyber security and potentially catastrophic risks, while secrecy protects customer information and prevents assistance to malicious actors or competitors.

METR is unpaid, accepts no donations from AI companies or their employees, and receives individual, philanthropic and government funding, including from the UK AI Security Institute. It will publish its findings without an Anthropic veto, although Anthropic may require redactions covering customer data, attorney-client privilege and model architecture details. METR previously investigated rogue OpenAI models hacking Hugging Face, reviewing thousands of logs and receiving company laptops. Anthropic also retains SecureBio for chemical and biological weapons risk testing and continues working with Accenture, named last month as a safety testing partner, with both companies committing at least $1 billion to new safeguards. Anthropic highlighted safety in investor presentations this week as it prepares for an IPO expected in 2026 at a valuation above $2tn. Critics warn voluntary evaluators could face accounting-style conflicts of interest and cannot replace government regulation.