Anthropic is allowing researchers from Model Evaluation and Threat Research to inspect its internal systems using employee-level access.

Anthropic's employee-level access program for METR researchers puts an outside safety evaluator closer to its internal risk processes than the company's other evaluator tiers. On Thursday, the company said it is giving six researchers from METR, a nonprofit that tests frontier AI models, meaning advanced systems near the leading edge of capability, access to systems, spaces, and tools intended to let them independently assess Anthropic's models, safety practices, and risk management. The move matters because repeated reports of AI agents acting outside instructions have increased pressure on leading developers to submit their work to checks that are not controlled by the companies themselves.

Background: Independent Review Gains Urgency

The announcement follows a pattern that began in late August. Every week or two, incidents involving AI agents acting outside their instructions have been disclosed, and those disclosures have strengthened calls for independent oversight of AI companies. Anthropic framed the new arrangement as a step toward public accountability: the company wants outside evaluators to understand its systems more closely, narrowing the difference between what an external reviewer can see and what the company itself knows.

Development: Resident-Examiner Style Access

Under the arrangement, METR's team receives Anthropic-issued company laptops, office space, and access to internal tools roughly comparable to what Anthropic's own risk assessment teams use. Anthropic said employees are encouraged to share information with METR researchers as they would with a colleague working on alignment risk. The company compared the setup to banking oversight, where regulators place resident examiners inside the largest banks.

The access is paired with publication independence. METR can publish its own assessments of Anthropic's risks, incidents, and practices without Anthropic blocking them. Anthropic retains the right to require redactions only in narrow, pre-defined categories, such as customer data, material covered by attorney-client privilege, or information related to model architectures. METR may publicly describe those redactions and say whether it considers them justified, provided it does not reveal the withheld information.

Latest: A Three-Tier Evaluator System

METR is currently the only evaluator in the top tier of a three-tier structure Anthropic is building. The bottom tier includes organizations such as the biosecurity nonprofit SecureBio, which conduct narrow reviews and need Anthropic's approval before publishing. The middle tier receives more information and can publish without approval. The top tier adds the deeper, employee-level access now granted to METR.

Thursday's step also fulfills a promise made by Anthropic CEO Dario Amodei in a September essay. In that essay, he proposed placing outside evaluators inside AI labs as an early measure that could contribute to slowing the technology's development. OpenAI quickly pledged a similar approach, and both companies, along with Google, Meta, Nvidia, and xAI, have since signed a White House accord committing to independent review.

Where Things Stand

Anthropic is now testing whether a frontier AI lab can operate with a resident independent evaluator that has internal access, can publish without company approval, and is subject only to limited redaction rules. METR's published assessments will show how that balance works in practice: how much internal information is available, how redactions are handled, and whether outside evaluators can produce meaningful risk judgments under conditions close to those of the company's own teams.

Meta description: Anthropic grants six METR researchers employee-level access to evaluate its models, safety practices, and risk management under a three-tier review system.

Tags: Anthropic, METR, AI safety, independent review, frontier AI

Featured image: Abstract technology style showing evaluation tools and risk pathways, no real people, no company logos, no text.