What happened
On 15 October 2026, Anthropic published “An update on independent evaluators”. It says Anthropic has started giving “employee-like access” to a team from METR, a nonprofit that evaluates frontier AI models. The post follows Anthropic’s earlier commitment to stronger third-party assessment. Anthropic makes Claude, which helped draft this article, so this is a company describing its own oversight arrangements. Anthropic’s page carries footnotes saying METR’s project terms were updated on 16 October and SecureBio’s on 15 October.
Key details
- METR: six METR employees receive company hardware, office space and internal tools and systems that Anthropic says are roughly on par with its own risk assessment teams. METR will publish its analysis of Anthropic’s risk levels, incidents or practices. Anthropic says it cannot veto publication but can require redaction of information in predefined categories such as customer data, privileged information and model architectures.
- Other partners: SecureBio is reviewing the chemical and biological weapons chapter of Anthropic’s August 2026 Risk Report. Accenture and Faculty are working toward a broader partnership on evaluation and red-teaming. Anthropic says early talks are underway with several other organizations, including Hugging Face.
- Three access tiers: (1) narrowly scoped investigations where Anthropic decides access and approves publication; (2) broader access and publication without Anthropic’s approval, though Anthropic can still require redactions and evaluators can say whether they think the redactions were justified; (3) the METR-style arrangement, with employee-like access and tier-2 publication rights.
- Safeguards: Anthropic says the METR team is small, works under enforceable agreements and is walled off from other METR staff, and that the arrangement is not a way for others to obtain information about Anthropic, its customers or its partners.
We have not seen METR’s own description of the arrangement; everything above is Anthropic’s account. Checked against Anthropic’s page on 18 October 2026.
Why it matters
Whether outside reviewers can see enough, and publish what they find, is a central question for judging AI company safety claims. This arrangement puts specifics on the table: who gets access, what they can publish, and what Anthropic can redact.
What to watch is the first METR publication. Until it appears, the practical effect is unknown, and the redaction categories, which Anthropic defines, determine how much the public will see.
Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article, and this article is about Anthropic. Statements are Anthropic’s own and have not been independently verified.
Sources
- An update on independent evaluators Primary source , Anthropic