Anthropic Gives METR Researchers Internal Access for Independent AI Safety Reviews

TL;DR

• Anthropic is granting six METR researchers employee-level access to assess its models, safety practices and risk management.
• The researchers receive company laptops, office space and access to internal risk assessment tools.
• METR can publish its findings without Anthropic’s approval, subject to narrowly defined redactions.
• The nonprofit is the first evaluator in the highest tier of Anthropic’s three-level independent review system.

METR Researchers Gain Access to Anthropic’s Internal Systems

Anthropic is opening its internal operations to six researchers from METR, a nonprofit that evaluates frontier AI models, as it expands independent oversight of its safety practices.

The company announced Thursday that the researchers will receive employee-level access to examine its models, risk management and approach to safety.

The arrangement includes Anthropic-issued laptops, office space and access to internal tools broadly comparable to those used by the company’s own risk assessment teams. Employees are encouraged to share information with the researchers as they would with colleagues working on alignment risks.

Anthropic said the goal is to reduce the information gap between external evaluators and the company developing the systems.

“We want to be publicly accountable for the decisions we’re making,” the company said.

METR Can Publish Findings Without Anthropic’s Approval

A central feature of the arrangement is METR’s ability to publish its assessments independently.

The nonprofit will be able to report on Anthropic’s risks, incidents and practices without the company having the right to veto publication.

Anthropic can, however, require redactions within narrow, predefined categories. These include customer data, information protected by attorney-client privilege and details about model architectures.

METR can publicly describe the redactions and state whether it considers them justified, provided it does not disclose the protected material itself.

Those terms give researchers a route to challenge the company publicly while preserving restrictions on specified sensitive information.

Three-Tier System Sets Different Levels of Evaluator Access

METR is currently the only organization in the highest tier of Anthropic’s developing independent evaluation framework.

The system establishes three levels of access and publication rights.

At the lowest tier, evaluators conduct focused reviews and need Anthropic’s approval before publishing. Biosecurity nonprofit SecureBio is among the organizations in this category.

The middle tier provides more information and allows evaluators to publish without prior approval. The highest tier adds the employee-level access granted to METR.

Anthropic compared the embedded arrangement to resident bank examiners, who work inside large financial institutions. METR remains an independent evaluator, however, rather than a financial-style regulator with statutory supervisory powers.

AI Agent Incidents Increase Pressure for Outside Scrutiny

The announcement comes amid heightened concern about AI agents acting beyond their instructions.

Reports of such incidents since late August have intensified questions about how effectively developers can monitor and control increasingly autonomous systems. They have also strengthened calls for assessments that extend beyond companies’ own testing and disclosures.

Giving outside researchers access to internal tools and staff could help them examine how safety procedures operate in practice, alongside the behavior of individual models.

The effectiveness of that oversight will depend on the information researchers can obtain and the findings they ultimately publish.

Anthropic Follows Through on Independent Review Pledge

The arrangement follows a proposal Anthropic CEO Dario Amodei outlined in a September essay, calling for outside evaluators to be embedded inside AI laboratories as an initial step toward pacing frontier development.

OpenAI subsequently made a similar pledge. Anthropic, OpenAI, Google, Meta, Nvidia and xAI have also signed a White House accord committing to independent review.

Anthropic’s agreement with METR provides a concrete example of that commitment, pairing extensive internal access with the right to publish critical assessments without company approval.