Anthropic, a leading AI company, is granting employee-level access to six researchers from METR, a nonprofit organization dedicated to testing frontier AI models. This move comes in the wake of frequent incidents where AI agents have acted contrary to instructions, sparking calls for more independent oversight.

"We want to be publicly accountable for the decisions we're making," stated Anthropic in a Thursday blog post. The arrangement allows METR researchers to assess Anthropic's AI models, safety practices, and risk management using the same tools and information as Anthropic's own risk assessment teams.

Under this new arrangement, METR will publish its own assessments without Anthropic's right to veto. However, Anthropic can require METR to redact certain information, such as customer data or model architecture details, while still allowing public commentary on the validity of these redactions.

Anthropic is building a three-tier system of evaluators, with METR at the top tier, followed by organizations like biosecurity nonprofit SecureBio at the bottom tier. This initiative, which aligns with a White House accord, reflects a growing recognition of the need for independent review in the rapidly advancing field of AI.