Anthropic Gives Six METR Researchers Inside Access to Evaluate AI Safety

Six researchers from the nonprofit METR now have employee-like access inside Anthropic. The arrangement was announced on October 15, 2026. Anthropic's October 16 clarification says it can require redactions of certain protected information and ultimately decides what falls within those categories.

Publication rights

Anthropic announced the arrangement with METR on October 15, 2026. Six researchers have employee-like access to risk assessment systems.

METR may publish analyses of Anthropic's risks, incidents and safety practices without obtaining the company's approval. But publication is not unrestricted. Anthropic can require redactions of customer data, attorney-client privileged information and details about model architectures. In an October 16 clarification, Anthropic said it ultimately decides what falls within those categories. METR can publicly describe the access it received and whether it considers redactions justified, without disclosing the protected material. This is a real publication right, but it does not eliminate the possibility of disputes about what can be shown.

How much internal access do the researchers have?

Anthropic says the six METR staff have company-issued hardware, office space and access to internal tools and systems roughly comparable to its own risk-assessment teams. Employees are encouraged to share relevant information with them, except for protected categories. The arrangement goes further than inviting an outside group to test a finished model through an API. Access to internal processes can help reviewers examine how failures were detected, how agent behavior was monitored and how risk decisions were made.

There are limits. Anthropic has not promised access to every model checkpoint or internal document. The embedded team is also separated from other METR employees under enforceable security agreements. The October announcement describes an access arrangement, not a completed independent finding that Claude is safe.

Anthropic's three tiers of independent evaluation

The company is building three levels of access. Tier 1 covers narrowly scoped reviews, with publication requiring Anthropic's approval. Tier 2 allows reviewers to publish without approval and provides more information, subject to defined redactions. Tier 3, currently exemplified by METR, adds employee-like access while retaining Tier 2-style publication rights.

Anthropic says it hopes to expand the highest tier. Other partnerships are not yet equivalent. SecureBio is reviewing the chemical and biological risks chapter of Anthropic's August Risk Report after an earlier pilot. Anthropic is also working toward a broader partnership with Accenture and Faculty covering model evaluations, red-teaming and safeguards. These efforts should not be described as six more embedded METR-style teams.

Why the timing matters

The October 15 announcement comes amid renewed scrutiny of AI agents behaving unexpectedly during evaluations. Earlier this month, Anthropic described cases in which agents took unintended actions on external websites and said it would restrict live internet access during internal evaluations while improving monitoring and containment.

METR and Anthropic had already worked together. In March, METR described a three-week exercise testing parts of Anthropic's internal agent-monitoring systems. The researchers identified vulnerabilities and produced test materials, but did not find evidence that severely undermined the major claims in the associated risk report. That earlier, limited review is not the same as the new continuing access agreement.

The point of deeper oversight is to examine more than a model's answers on a benchmark. Investigators also need to understand how a lab noticed failures and what it did about them.

What independent oversight can and cannot prove

Anthropic compares embedded evaluators to resident examiners at major banks. The analogy captures the need for an outside party to see information that public disclosures miss, but it is not a legal equivalence. METR is a nonprofit research organization, not a regulator with statutory enforcement powers.

Open questions remain: which internal evidence reviewers can inspect in practice, how disagreements over redactions will be resolved, how often METR will publish and whether other AI developers will provide similar access. Businesses deploying AI agents should not treat this arrangement as an independent certification of Claude.

Belliste's assessment

The meaningful commitment is continuing access combined with the right to publish criticism. The limitation is Anthropic's continuing control over protected information. The arrangement will earn credibility through detailed published findings and transparent handling of disagreements, not through the access announcement alone.

Editorial scope: This is a report on an oversight agreement, not a completed independent safety audit.

The underlying documents are Anthropic's October 15 announcement, updated October 16, and METR's March report on Anthropic's agent-monitoring systems.

Hero photo: Kevin Ache / Unsplash. Illustrative server-room photograph; it does not show an Anthropic or METR facility.