Anthropic says it is giving employee-like access to METR as part of its pledge to partner with evaluators for assessment of its models and safety practices (Anthropic)

In a recent announcement, Anthropic revealed its decision to grant employee-like access to a team from METR, a nonprofit organization specializing in assessing frontier AI models. This move comes as part of Anthropic's pledge to collaborate with independent evaluators for assessing its models and safety practices. Earlier this year, Anthropic signed the White House Accord on Super Intelligence, committing to such partnerships.

METR will publish the results of its risk assessments, allowing for transparency in Anthropic's safety work. While Anthropic cannot publicly share all information relevant to its safety efforts due to potential misuse by malicious actors or commercial sensitivity of certain data, such as compute or training information, it aims to bridge the gap between what evaluators understand about Anthropic's systems and what the company itself comprehends.

Anthropic employees are encouraged to share information with METR just as they would with colleagues working on alignment risk, with certain exceptions, like data about customers, attorney-client privileged information, or details about Anthropic's model architectures. METR will have access to Anthropic's internal tools and systems, similar to those used by Anthropic's own risk assessment teams.

This collaboration sets a new standard for safety oversight in the AI industry, as Anthropic moves forward with multiple evaluators and actively learns about the dynamics of such partnerships. Anthropic is committed to sharing updates on its engagements with evaluators. The company has onboarded six METR employees with company hardware, office space, and access to internal tools and systems, similar to those of Anthropic's risk assessment teams.

However, Anthropic employees are restricted to collaborating with a small team operating under enforceable agreements, while other METR employees remain siloed from these activities.

This approach to embedded evaluation provides evaluators with a close view of Anthropic's operations, ensuring that they can publish their findings with Anthropic's approval. As these partnerships progress, Anthropic aims to embed more independent evaluators at higher tiers of access, with the ultimate goal of achieving publication rights for all evaluators, similar to the third tier arrangement with METR.