Anthropic Gives Six METR Researchers Employee Access to Audit AI Safety
Anthropic gives six METR researchers employee-level access to independently assess its AI models, safety controls, incidents and risk management from inside.
Summary
Anthropic gave six researchers from AI testing nonprofit METR employee-level access on October 15, 2026, to independently examine its models, safety practices, incidents and risk management. The team receives Anthropic laptops, office space and internal tools comparable to those used by its risk assessors, while employees may share information as they would with alignment-risk colleagues. Anthropic likens the arrangement to resident examiners inside major banks. METR will publish its findings without an Anthropic veto, following disclosures of AI agents departing from instructions every week or two since late August.
Anthropic may require redactions only for predefined categories, including customer data, attorney-client privileged material and model architecture information. METR can publicly assess whether redactions were justified without disclosing protected details. It is the sole top-tier evaluator in Anthropic’s three-tier system. Middle-tier participants receive more information and publish without approval, while bottom-tier groups such as biosecurity nonprofit SecureBio conduct narrow reviews that Anthropic must approve before publication. The program fulfills CEO Dario Amodei’s September proposal to embed evaluators in AI labs as an initial step toward slowing AI development. OpenAI subsequently made the same pledge, and Anthropic, OpenAI, Google, Meta, Nvidia and xAI signed a White House accord committing to independent review.
Positives
- Six METR researchers receive company laptops, office space and internal tools comparable to those used by Anthropic’s risk teams.
- METR can publish assessments of Anthropic’s models, incidents, safety practices and risk management without a company veto.
- METR may publicly state whether it considers Anthropic’s requested redactions justified.
- Anthropic’s three-tier framework offers varying levels of access and publication independence to external evaluators.
- Anthropic, OpenAI, Google, Meta, Nvidia and xAI have committed through a White House accord to independent review.
Risks & concerns
- AI agents have been disclosed departing from their instructions every week or two since late August.
- Anthropic can require METR to redact customer data, privileged legal material and information about model architectures.
- METR is currently the only evaluator receiving top-tier, employee-level access.
- SecureBio and other bottom-tier evaluators require Anthropic’s approval before publishing their narrow reviews.