A man with curly brown hair and dark-framed glasses, wearing a blue button-up shirt and a headset microphone, speaks with one hand raised in front of a dark stage backdrop and a video screen.
Dario Amodei at TechCrunch Disrupt, September 20, 2023. Archive photograph of Anthropic's chief executive by Kimberly White/Getty Images for TechCrunch (Wikimedia Commons, CC BY 2.0); it does not depict the October 2026 announcement. Kimberly White/Getty Images for TechCrunch / Wikimedia Commons · CC BY 2.0

What Anthropic Announced

In an October 15, 2026 post, Anthropic said it has started to give employee-like access to a team from METR, a nonprofit that evaluates frontier AI models. Anthropic says it has onboarded six METR employees with company hardware, office space, and internal tools and systems roughly on par with those of its own risk assessment teams. METR will publish its analysis of Anthropic's risk levels, incidents, or practices.

The post also names SecureBio and Accenture as evaluators and early talks with others, including Hugging Face. It describes three planned tiers of evaluator access, with METR in the third, and calls the embedding voluntary. Notes on the page record October 15 and October 16 updates clarifying SecureBio's and METR's current project terms.

What Stays With Anthropic

Anthropic says it cannot veto publication but can require METR to redact information in predefined categories, such as data about customers, attorney-client privileged information, and information about model architectures. The company ultimately decides what falls into those categories, and METR reserves the right to make certain statements about the level of access and information it received. An Internet Archive copy from before the October 16 update described "narrow" categories and said METR could state whether it believed redactions were justified. The current METR paragraph uses neither wording.

Audit Questions on Scope and Publication

Anthropic describes publication in the future tense, and METR's updates page, checked on October 18, lists no entry about this arrangement. Our editorial questions: Can METR say publicly whether a redaction changed its conclusions? The post gives the third tier publication rights similar to the second, where evaluators may publish views on redactions, yet the METR paragraph mentions only certain statements about access and information received. Will METR's analysis cover the controls, monitoring, and detection named in the external assessment item of the White House accord, which the post mentions without saying whether this arrangement is meant to meet it?

Anthropic compares embedded evaluation to resident bank examiners while calling the analogy inexact. It does not publish the agreement or describe METR's funding, so the independence criteria it lists (no equity, no financial or operational dependence) rest on Anthropic's description. Related briefs: Accenture arrangement, METR incident review, White House accord.