The key facts

  • Six METR researchers received employee-level access at Anthropic for independent evaluation of models and risks
  • METR can publish evaluations without Anthropic's approval; there is no veto, but redactions are possible in narrow categories
  • This is the top tier of a three-tier system of external evaluators; the promise was made in Dario Amodei's September essay

Employee-level access for external researchers

Anthropic has granted six researchers from the nonprofit METR, which tests frontier AI models, employee-level access to independently evaluate the company's models, safety practices, and risk management. The company announced this in a blog post on Thursday, October 15, 2026. METR received corporate laptops, office space, and access to internal tools from Anthropic at roughly the level used by the company's own risk evaluation teams. Anthropic employees are encouraged to share information with the researchers the same way they do with colleagues working on alignment risk.

Context: increasing incidents involving AI agents

As TechCrunch notes, the agreement came amid a growing number of incidents: since late August, cases where AI agents acted outside their given instructions have surfaced roughly every week or two. This has intensified calls for independent oversight of AI companies. Anthropic stated: 'We want to be publicly accountable for the decisions we make' and aims to 'reduce the gap between what evaluators understand about the frontier AI company's systems and what the company itself understands.' The company compared the arrangement to banking practice, where regulators place resident examiners inside the largest banks.

Right to publish without company veto

METR will publish its own evaluations of Anthropic's risks, incidents, and practices, and the company has given up the right to block publications. According to Anthropic, 'Anthropic cannot veto a publication, but may request redaction of information if it falls into narrow, pre-defined categories' — for example, customer data, materials under attorney-client privilege, or information about model architectures. METR is entitled to publicly characterize these redactions and comment on their justification without disclosing the classified information.

Three-tier system of evaluators

METR is so far the only evaluator at the top of the three-tier system Anthropic is building. Lower-tier evaluators, such as the biosecurity nonprofit SecureBio, conduct narrow checks and need Anthropic's approval to publish. The middle tier receives more information and may publish results without approval, while the top tier adds employee-level access, as METR has.

Fulfilling Amodei's September promise

Thursday's announcement fulfills a promise that Anthropic CEO Dario Amodei made in a September essay, in which he proposed embedding external evaluators into AI labs as a first step toward slowing the development of the technology. OpenAI quickly promised to do the same, and both companies, along with Google, Meta, Nvidia, and xAI, have since signed an agreement with the White House committing to independent oversight.