Anthropic has granted unprecedented access to its internal systems for third-party AI safety evaluators. Six employees from METR, a nonprofit organization focused on assessing frontier AI models, have received "employee-like access," which includes company hardware, office space, and entry to internal tools and systems comparable to those used by Anthropic's own risk assessment teams. This arrangement marks a significant step in the AI industry's ongoing efforts to enhance safety and transparency through independent oversight.
The initiative stems from Anthropic's commitment, announced in September, to strengthen independent, third-party assessments of its models and safety protocols. The company also signed the White House Accord on Super Intelligence, a pledge by AI firms to collaborate with external auditors. Anthropic stated that it cannot publicly disclose all information related to its safety work due to potential misuse by malicious actors or commercial sensitivities. By providing deep access to METR, Anthropic aims to bridge the gap between what external evaluators understand about the company's systems and what the company itself knows.
Under the agreement, METR will publish its findings on Anthropic's risk levels, incidents, and practices. Anthropic retains the right to request redactions for specific categories of information, including customer data, attorney-client privileged information, and details about model architectures. However, METR is permitted to publicly comment on the justification for any redactions. This tiered access system, with METR operating at the highest level, is part of Anthropic's broader strategy for engaging with external evaluators, which also includes partnerships with SecureBio, Accenture, and Faculty.
The move follows recent regulatory pressure on AI labs, particularly after security incidents prompted government action. Last week, the government began requiring AI companies to disclose safety incidents and security breaches following Anthropic's own report of an incident involving its Claude model. Anthropic's CEO, Dario Amodei, previously proposed embedding evaluators from nonprofits like METR as a crucial step for pacing the development of advanced AI capabilities.
The criteria for effective AI evaluators, as outlined by Anthropic, include expertise in assessing frontier model capabilities, independence from financial ties to the company being evaluated, sufficient access to assess risks, and freedom to publish their findings. This approach to external auditing is seen as a voluntary alternative to government-mandated oversight, which is still developing in the United States.