Anthropic has started giving a team from METR, the nonprofit that evaluates frontier AI models, "employee-like access" to its internal systems — company hardware, office space, and tools roughly on par with what its own risk assessment teams use — so the external group can independently assess the risks of Anthropic's frontier models and publish what it finds.
The arrangement, announced on Anthropic's blog on Wednesday, goes further than the audit-style relationships AI companies have typically tolerated, and it comes with a provision that stands out: METR will publish the results of its risk assessments, and Anthropic cannot veto publication.
What "Employee-Like Access" Actually Means
Six METR employees have been onboarded with company hardware, office space, and access to internal tools and systems that Anthropic describes as roughly on par with what its own risk assessment teams receive. Outside a small number of clearly defined categories — data about customers, attorney-client privileged information, and information about model architectures — Anthropic says its employees are encouraged to share information with the METR team just as they would with a fellow employee working on alignment risk.
The engagement is deliberately narrow. It covers a small team operating under enforceable agreements, and other METR employees are siloed from that team's work — including researchers who collaborate with Anthropic on other projects. Anthropic notes the setup gives a few qualified people, working under Anthropic's security standards, a close view of how the company operates; it is not a channel for other organizations to obtain Anthropic's information or that of its customers.
Publication Rights, With Guardrails
Under the agreement, METR will publish its analysis of Anthropic's risk levels, incidents, or practices. Anthropic cannot block publication, but it can require redactions when information falls into the predefined categories described above — and Anthropic itself decides what does or does not fall into those categories. METR, for its part, reserves the right to make certain statements about the level of access and information it received, a counterweight designed to keep the redaction power from becoming a mute button.
Why Anthropic Is Opening Up
The move follows a commitment Anthropic made last month to stronger independent, third-party assessment of its models and safety practices, and its signature on the White House Accord on Super Intelligence, in which AI companies pledged to partner with independent external auditors or evaluators. It also extends an existing relationship: over the past year, Anthropic worked with METR on several reviews of its Risk Reports.
METR is not alone in the portfolio. Anthropic says it is also working with SecureBio and Accenture as evaluators — Accenture's Faculty was named the company's first embedded evaluator in September — and is in early discussions with several other organizations, including Hugging Face. Those earlier-stage partnerships come with more limited information access and publication rights while the company learns what kinds of partnership work.
The Bank-Supervision Analogy
Anthropic's own framing borrows from finance. Banks know more about their own risk positions than outsiders do, so regulators assign resident examiners to work inside the largest banks; if a bank's risk-taking could threaten the public, third parties try to see that in advance. The company positions embedded evaluation as the AI industry's version of that logic — a step toward robust safety oversight in a field where the builder always knows more than the watchdog.
Underlying it all is a stated goal: closing the gap between what independent evaluators understand about a frontier AI company's systems and what the company itself understands.
What Evaluators Need
The announcement also lays out what Anthropic thinks meaningful evaluation requires: expertise in assessing frontier model capabilities, including understanding how evaluations can go wrong through phenomena like alignment faking; independence, meaning no equity in or financial or operational dependence on the company being evaluated; enough access to meaningfully evaluate risks, extending beyond the latest models to how they were trained and how safeguards function; and freedom to publish observations and assessments of a model or a process.
A Test Case for the Whole Industry
The arrangement will be judged by its output. METR's first published assessment under the new access regime will show whether employee-like access produces findings with real teeth or reports trimmed by redaction categories that the company itself defines. Skeptics will note that Anthropic controls the category boundaries and that the pilot team is small; supporters will note that a publication right Anthropic cannot veto is genuinely novel among frontier labs.
Either way, the industry is watching. If METR's embedded access surfaces material findings and publishes them unimpaired, the model becomes the obvious template for every lab that signed the White House accord — and a benchmark for policymakers weighing whether voluntary oversight can substitute for mandated audits. For ongoing coverage of AI safety policy and the AI industry news that shapes it, AI Buzz Wire follows each development as it breaks.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
