Anthropic Grants METR Researchers Employee-Level Access for Independent AI Safety Reviews

Anthropic is giving six researchers from METR, an independent nonprofit focused on evaluating advanced AI systems, employee-level access to examine its models, safety procedures and internal risk management. The arrangement gives outside researchers unusually deep visibility into how the AI company assesses potential dangers from its technology.

In an announcement on October 15, Anthropic said the researchers had received company-issued laptops, office space and access to internal tools comparable to those used by its own risk assessment teams. Employees are encouraged to share relevant information with the researchers as they would with internal colleagues working on AI alignment risks.

The initiative comes amid growing scrutiny of increasingly autonomous AI agents and calls for independent oversight of frontier AI developers.

METR Can Publish Findings Without Anthropic’s Approval

Under the agreement, METR will independently assess Anthropic’s risk levels, safety incidents and internal practices. The nonprofit can publish its conclusions without requiring the company’s approval, a significant provision intended to strengthen the credibility of external evaluations.

Anthropic retains limited authority to require redactions involving customer information, attorney-client privileged material and sensitive details about model architectures. However, METR can publicly explain whether it considers those redactions justified, as long as it does not disclose protected information.

Anthropic compared the arrangement to financial regulators placing examiners inside major banks, allowing independent specialists to evaluate risks that outsiders cannot easily observe.

Anthropic Introduces Three Levels of Independent Evaluation

METR currently occupies the highest tier of Anthropic’s proposed three-level evaluation framework. Researchers at this level receive employee-like access and independent publication rights.

At the lowest tier, organisations conduct narrowly defined reviews and need Anthropic’s approval before publishing findings. The middle tier provides broader access and publication independence but stops short of the internal access granted to METR.

Biosecurity nonprofit SecureBio operates in the first tier, while Anthropic is also developing evaluation partnerships with Accenture and other organisations.

The announcement follows a September essay by Anthropic CEO Dario Amodei advocating stronger independent assessments and greater caution as AI capabilities advance. Anthropic has also signed the White House Accord on Super Intelligence, which includes commitments to independent external evaluations.

The embedded research program marks a shift toward allowing outside experts to examine AI safety decisions from within a leading developer, rather than relying solely on public reports and limited model testing.