Anthropic grants nonprofit evaluator embedded access to assess AI safety risks
METR, a third-party watchdog, now has an on-site team with access to Anthropic's hardware and internal systems to evaluate the safety of its Claude models.
What happened
Anthropic announced Thursday that it has given METR, a nonprofit evaluator that has also worked with OpenAI, Google, and Meta, "employee-like access" to its offices, hardware, and internal systems. METR has embedded a team at Anthropic to assess risks posed by the company's AI models. This move fulfills a commitment Anthropic CEO Dario Amodei proposed in September, which garnered rare support from OpenAI's Sam Altman and Elon Musk, to place external auditors in leading AI labs. Anthropic framed the arrangement as necessary because stronger safeguards are required as models become more powerful, while noting it cannot publicly share all information relevant to its safety work without benefiting malicious actors or competitors.
Context
Third-party evaluators embedded in AI labs represent a concrete step toward addressing rising concerns about the risks posed by increasingly capable AI systems to the economy, cybersecurity, and broader society. The arrangement is designed to provide public accountability while protecting proprietary and sensitive safety information. Anthropic compared the model to bank supervision, where regulators assign resident examiners to major financial institutions. The announcement marks the first implementation of the pledges made by leading AI companies in response to safety concerns, though OpenAI has not yet laid out plans to bring in evaluators.
What's disputed
Critics have argued that embedded evaluator arrangements could lead to conflicts of interest, citing parallels to issues in the accounting sector, and have contended that greater government regulation is required rather than industry self-regulation through third-party evaluators.