Anthropic says six METR evaluators get employee-like access

Anthropic said it has onboarded six METR employees with company hardware, office space and access to internal tools and systems broadly comparable to what its own risk-assessment teams receive. The arrangement gives the nonprofit evaluator a close view of Anthropic’s safety work while excluding specified categories of information.

The move implements a commitment CEO Dario Amodei made in September to embed independent evaluators with ongoing, employee-like access. Anthropic describes METR’s placement as the third and deepest tier in a three-tier evaluator framework. Earlier-stage partnerships have more limited access and publication rights. The external arrangement is separate from Anthropic’s internal evaluation controls.

The access has defined limits. Anthropic said METR will not receive customer data, attorney-client-privileged information or details about model architectures. Architecture information may be shared when highly relevant, but only with explicit internal approval.

Anthropic also said the work is confined to a small team operating under enforceable agreements. Other METR employees, including researchers working on separate Anthropic projects, are siloed from the embedded team.

Anthropic said it cannot veto METR publications about risk levels, incidents or company practices. It may require redactions only within narrow, predefined categories. METR can publicly characterize those redactions and say whether it considers them justified, provided it does not disclose the withheld information.

Anthropic describes the arrangement as voluntary and has urged other frontier AI developers to adopt the practice. In a proposal published one day before Anthropic’s update, METR said voluntary assessments ultimately depend on companies granting access and called for reports to disclose their scope, access, redactions and other limitations. METR said company-discretionary access means its embedded assessors would not have the same access as senior internal risk staff.