Anthropic said in a blog post on Thursday that it is granting employee-level access to six researchers from METR, a nonprofit organization that evaluates frontier AI models. The access is intended to let METR independently assess Anthropic’s models, safety practices, and risk management.
The move comes amid a series of incidents disclosed since late August in which AI agents have acted beyond their instructions, occurrences that have increased calls for external, independent scrutiny of AI companies. Anthropic wrote that it wants to be publicly accountable for its decisions and to reduce the information gap between what external evaluators know about a frontier AI company’s systems and what the company itself understands.
Scope of METR’s access
Under the arrangement, the METR team will receive Anthropic-issued company laptops, office space, and access to internal tools roughly equivalent to those used by Anthropic’s own risk assessment teams. Anthropic said its employees are encouraged to share information with the METR researchers as they would with any colleague working on alignment risk.
Anthropic likened the setup to banking oversight, comparing METR’s presence to resident examiners that regulators place inside the largest banks.
Publication and redaction rules
METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and Anthropic has relinquished the right to veto those publications. The company can, however, require redaction of information that falls into narrow, predefined categories — for example, customer data, attorney-client privileged material, or details about model architectures. Anthropic said METR is free to describe any redactions publicly and state whether it believes they were justified, as long as the redacted content itself is not revealed.
Three-tier evaluation system
METR is currently the only evaluator placed at the top tier of Anthropic’s three-tier evaluation system. Lower-tier evaluators, such as the biosecurity nonprofit SecureBio, conduct narrow reviews and must obtain Anthropic’s approval to publish. Middle-tier evaluators receive more information and can publish without Anthropic’s approval, while the top tier adds the employee-level access provided to METR.
Context and related commitments
Thursday’s announcement fulfills a pledge made by Anthropic CEO Dario Amodei in a September essay proposing that embedding outside evaluators inside AI labs could be a first step toward slowing the technology’s development. OpenAI made a similar pledge soon afterward. Anthropic, OpenAI, and the companies Google, Meta, Nvidia, and xAI have since signed a White House accord committing to independent review.
