Skip to main content

Anthropic gives METR employee-like access for safety evaluation

17 OCTOBER 2026·2 MIN READ·4 SOURCES

Anthropic said on Oct. 15, 2026, that it has started giving a team from nonprofit METR employee-like access to its operations, following a commitment to stronger independent, third-party assessments of its models and safety practices. Anthropic said it onboarded six METR employees and provided them with company hardware, office space and internal systems access.

Anthropic gives METR employee-like access for safety evaluation

Key takeaways · 4

  • 01

    The METR team includes six employees and has company hardware, office space and internal systems access.

  • 02

    METR may publish analyses of Anthropic’s risk levels, incidents and practices without an Anthropic veto.

  • 03

    Anthropic can require redactions for predefined categories and makes the final call on whether information fits those categories.

  • 04

    Anthropic places METR’s arrangement in the highest of its planned three evaluator-access tiers.

Access is designed to resemble internal work

Anthropic said the METR team’s access to its internal tools and systems is roughly comparable to access available to Anthropic’s risk assessment teams.[1] Anthropic also said employees are encouraged to share information with METR as they would with a colleague working on alignment risk.[1]

Publication rights come with redaction limits

Anthropic said METR may publish analyses of its risk levels, incidents and practices without requiring Anthropic’s approval.[1] Anthropic may nevertheless require METR to redact information in predefined categories, including customer data, attorney-client privileged information and model architecture information.[1] Anthropic said it ultimately decides whether information belongs in those categories.[1] The terms therefore allow publication while leaving Anthropic a defined role in restricting specific content.

Anthropic places METR in its third tier

Anthropic said it is developing a three-tier system for evaluator access and hopes to have several evaluators at its highest tier over time.[1] In its description, first-tier evaluators conduct narrowly scoped investigations and need Anthropic’s approval to publish; second-tier evaluators receive more information and may publish without approval, although Anthropic can still require specific information to be removed.[1] Anthropic said METR’s employee-like access and publication rights put this engagement in the third tier.[1] The tier description offers a framework for comparing access and publication terms, but Anthropic did not say that several evaluators already operate at the highest tier.

The engagement sits within broader evaluation work

Anthropic said it had worked with METR over the previous year on several reviews of its Risk Reports.[1] It also said SecureBio was reviewing the chemical and biological weapons chapter of its August Risk Report after publishing a pilot review of an earlier report.[1] Anthropic described work toward a partnership with Accenture and Faculty covering model evaluation and red-teaming, alignment assessments, safeguard testing and capability-building, and said discussions with several other organizations, including Hugging Face, were at an early stage.[1] METR says voluntary company engagements leave it dependent on companies for participation, access and permission to publish.[2]

For teams assessing AI risks, the arrangement makes evaluator access and publication terms part of the practical design of an independent review. Professionals can use those terms to judge what an evaluation may reveal, what could remain undisclosed and where the company retains discretion.

Why it matters
Story quiz

Test yourself on this story — 1 question.

Create a free account to take the quiz, earn XP, and get a daily session built for your industry.

Take the quiz

How this developed

  1. 17 October 2026

    Anthropic gives METR employee-like access for safety evaluation

Sources

AI fluency, one session a day, built for your work.