Anthropic gives nonprofit AI evaluator METR employee-like access to assess Claude model risks

Anthropic has granted METR, an external nonprofit evaluator, “employee-like access” to its hardware and internal systems to assess risks from its AI models. The move follows Anthropic’s commitment to increase safeguards amid concerns that more capable AI systems could pose serious risks to cyber security, the economy, and broader society.

For AI builders, Anthropic’s move shows one concrete way to support third-party evaluations: deeper access can produce more informed risk assessments, but it also requires sharing sensitive internal systems. The article does not say what METR can publish, how independent its findings will be, or whether this access changes model release decisions.