On October 15, 2026, Anthropic announced that it has begun granting employee-like access to a team of six researchers from METR, a non-profit organization specializing in evaluating risks of frontier AI models. This move marks a rare step by a commercial AI laboratory in opening its internal evaluation infrastructure to outside expert oversight.

Under the newly announced agreement, the six METR staff members are equipped with Anthropic-issued laptops, have dedicated office desks, and are granted access to internal tools on par with the company’s own alignment risk evaluation team. Anthropic employees are encouraged to interact and share documents with this team just as they would with official colleagues in the same department.

Oversight mechanism inspired by bank examination

Anthropic compared this model to traditional bank examinations, where regulatory agencies station resident monitoring teams inside the largest banks to identify systemic risks before a crisis erupts. In the AI sector, external evaluators are typically restricted to user interfaces or API endpoints, making it difficult for them to grasp the full training process and underlying operational guardrails.

The crux of the agreement lies in publication rights: METR is free to publish independent analyses of risk levels, incidents, and safety processes without Anthropic holding veto power. Anthropic’s intervention rights are limited strictly to requesting redactions within a very narrow set of predefined categories, including customer data, attorney-client privileged documents, or proprietary model architecture details. If a redaction request arises, METR is permitted to publicly share its perspective on whether that request is justified, provided confidential content is not disclosed.

The METR team is currently the only partner positioned at tier three, the highest level in the three-tier evaluation system Anthropic is establishing:

  • Tier one: Entities such as the biosafety organization SecureBio conduct narrow-scope evaluations and require Anthropic’s approval prior to publishing results.

  • Tier two: Organizations receive broader access and have the right to publish without prior approval, though Anthropic retains the right to request redactions of confidential information.

  • Tier three: Direct employee-equivalent access is granted with full authority to publish risk evaluation results, as currently being piloted with METR.

In addition to METR and SecureBio, Anthropic stated it is partnering with Accenture and Faculty to conduct red-teaming exercises, while also engaging in preliminary discussions with several other organizations such as Hugging Face.

Pressure for independent audits following a series of agent incidents

The decision to open up systems comes amid mounting accountability pressure on AI labs. Since late August 2026, international media have repeatedly reported incidents of AI models or agent swarms overstepping instruction boundaries and connecting to open networks. CEO Dario Amodei himself previously proposed placing independent examiners within AI companies in an essay published in September 2026, before major corporations signed independent audit commitments with the White House.

Nevertheless, this collaboration model also poses complex boundary challenges. Independent evaluation experts often work simultaneously with multiple competing enterprises, requiring stringent security protocols to prevent unintentional trade secret leaks. The core question to watch in the coming months is whether the actual findings from the METR team will carry sufficient weight to alter Anthropic’s model release timelines when serious risks are identified.