Anthropic gives six METR evaluators employee-like access to its AI safety work

• Anthropic says six METR employees have company hardware, office space and access to internal systems broadly comparable to those used by its risk-assessment teams. ^[1]
• Anthropic says METR can publish assessments of the company’s risk levels, incidents and practices without Anthropic’s approval. ^[1]
• Anthropic can require redactions for predefined categories including customer data, attorney-client material and model-architecture information. ^[1]
• The company has created three evaluator tiers; first-tier partners need Anthropic’s approval to publish, while higher-tier evaluators receive broader access and publication rights. ^[1]
• Public materials reviewed for this article do not specify the METR agreement’s funding, duration, termination provisions or first-publication date. ^[1]^[2]

Anthropic says it has begun giving six employees of nonprofit evaluator Model Evaluation and Threat Research, or METR, access to company hardware, office space and internal tools used by its risk-assessment teams. The company says the evaluators will examine its models, safeguards and internal practices and publish their findings without Anthropic’s approval. [1]

The arrangement expands on METR’s previous work reviewing Anthropic risk reports and conducting red-team exercises. It gives the nonprofit closer access to the company’s internal safety work, while preserving limits around customer information, attorney-client privilege and sensitive model-architecture details. [1][3]

Access with defined boundaries #

Anthropic says METR’s team will receive access “roughly on par” with its internal risk-assessment teams. Employees are encouraged to share information with the embedded evaluators as they would with colleagues working on alignment risk, subject to a small number of exclusions. [1]

The stated remit covers Anthropic’s risk levels, incidents and practices. The October 15 announcement, which Anthropic updated October 16 to clarify METR’s terms, does not say that the team can halt a model release, direct engineering work or compel a particular safety measure. [1]

Anthropic retains the right to require redactions when information falls into predefined categories. The company names customer data, attorney-client privileged information and model architectures as examples. It says METR can publish statements about the level of access and information it received, including its views on whether a redaction was justified. [1]

Anthropic’s public announcement does not state who funds the METR engagement, how long it lasts, what termination protections apply or when METR will publish its first assessment. Those terms remain material to judging the arrangement’s independence and durability. [1][2]

A publication rule different from earlier reviews #

Anthropic’s framework separates the METR arrangement from its first-tier evaluator relationships. First-tier evaluators conduct narrowly scoped investigations, with Anthropic deciding the access provided and approving publication. At the second tier, evaluators can publish without approval but receive more limited access than METR. The company places METR’s arrangement in a third tier with employee-like access and publication rights similar to the second tier. [1]

That is a meaningful change from Anthropic’s earlier safety-review structure, which centered on company-authored risk reports and external feedback. Anthropic’s Responsible Scaling Policy, first published in September 2023, established a voluntary framework for assessing catastrophic risks and applying safeguards. ^[4] Its February 2026 update said third-party reviewers would receive unredacted or minimally redacted access to risk reports and publicly review the company’s reasoning and decisions in specified circumstances. [5]

The new arrangement extends review beyond selected documents to parts of the organization that produces them. It does not, however, eliminate company control over access or protected information. Anthropic decides what access to provide lower-tier evaluators and whether disputed material falls within its redaction categories. [1]

METR’s prior work offers a limited baseline #

METR has already worked with Anthropic on external reviews. In March, the nonprofit said one of its staff members spent three weeks red-teaming a subset of Anthropic’s internal agent-monitoring and security systems. METR later incorporated findings from that exercise into a broader Frontier Risk Report involving Anthropic, Google, Meta and OpenAI. [3][6]

That history shows that outside testing has produced public material, but it does not independently establish the terms of the new embedded arrangement. Anthropic is the source for the six-person staffing, access level, redaction rules and publication rights described in the October announcement. [1]

Anthropic also says it is working with SecureBio and Accenture and discussing possible projects with other organizations, including Hugging Face. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic’s August risk report; the company describes its other evaluator relationships as less mature and more limited in access or publication rights. [1]

The broader model remains voluntary. The Associated Press reported that experts welcomed embedded review while questioning how evaluators would be selected and how standards would be set. Brookings fellow Elham Tabassi characterized company-provided audits as voluntary and company-controlled until their terms are made public. [7]

The first METR publication under the new terms will be the clearest test of the arrangement. It should show whether the evaluator can report material criticism, describe meaningful access limits and explain any redactions without Anthropic’s approval becoming an indirect constraint.

Companies mentioned #

Further sources #

The stories that matter, in one email. Free — unsubscribe anytime.