METR’s team inside Anthropic can publish without a veto

Six METR employees now work inside Anthropic. They have company hardware, office space and access to internal tools and systems roughly on par with Anthropic’s own risk assessment teams, the company said in an October 15 post from Anthropic. METR, a nonprofit that evaluates frontier models, will publish its analysis of Anthropic’s risk levels, incidents or practices, and Anthropic can’t veto it.
METR isn’t new to Anthropic. Over the past year it reviewed several of Anthropic’s Risk Reports. What’s new is where it sits. Before, it read documents Anthropic had already written. Now its people work alongside staff who are told to share information with them “just as they would with a fellow employee working on alignment risk.” The move follows two commitments from September: stronger third-party assessment of Anthropic’s models and safety practices, and Anthropic’s signature on the White House Accord on Super Intelligence, where companies pledged to partner with independent external auditors or evaluators.
A few categories stay off limits: customer data, attorney-client privileged information and details of model architectures. Anthropic can also require METR to redact material in those narrow, predefined categories before publishing. METR may say publicly how much was redacted and whether it thinks the redactions were justified, as long as it doesn’t reveal what was removed. But Anthropic still decides, in the end, what falls into those categories.
The arrangement has its own walls. The embedded team is small and works under enforceable agreements. Other METR staff are siloed from its work, including researchers who work with Anthropic on other projects. “It’s not a way for other organizations to obtain our information, or that of our customers and partners,” the post says.
Anthropic’s case for all this comes down to what it can’t publish. Some safety-relevant material could help malicious actors, and some is commercially sensitive, such as compute or training information. So instead of publishing it, Anthropic gives outside evaluators deep access and has them publish evidence and reasoned conclusions. The stated goal is to close the gap between what evaluators understand about a frontier lab’s systems and what the lab itself understands. Anthropic compares the setup to bank supervision, where regulators place resident examiners inside the largest banks because banks know their own risk positions better than outsiders do. It adds that the analogy isn’t exact.
Anthropic sets four requirements for evaluators. The first is expertise in testing frontier models, including knowing how evaluations go wrong. Its example is alignment faking, where a model behaves well because it believes it is being watched. The second is independence: no equity in the company and no financial or operational dependence on it. The third is enough access, which means the latest models plus visibility into how they were trained and how safeguards work. The fourth is the freedom to publish.
Not every evaluator gets METR’s deal. Anthropic is building three tiers of access:
- First tier: narrowly scoped risk investigations. Anthropic decides what access to give, and findings are published with its approval.
- Second tier: more information, and publication without Anthropic’s approval. Anthropic can still require specific categories to be removed, and the evaluator can say publicly whether it thinks the redactions were justified.
- Third tier: employee-like access, with publication rights similar to the second tier. METR is the only evaluator here so far.
SecureBio is in the first tier. It completed and published a pilot review of Anthropic’s previous risk report, and is now reviewing the chemical and biological weapons chapter of the August Risk Report. Once that’s done, Anthropic says it will consider projects with more access and publication rights.
Accenture and Faculty are working with Anthropic on a broader partnership: evaluating and red-teaming models, running alignment assessments, testing safeguards and building capabilities across the organizations. In September, Anthropic and Accenture said they expect to invest at least $1 billion over five years in this area. Anthropic is also in early discussions with several other organizations, Hugging Face among them. For now, every evaluator except METR gets more limited information and narrower publication rights. Anthropic says the early engagements are meant to check that partners protect confidential information and have the resources and expertise to evaluate its models. Over time, it hopes to have several evaluators at the third tier, each publishing its own analysis of Anthropic’s risk posture and key decisions.
“We’re embedding evaluators voluntarily, and we believe every frontier AI company should do the same,” the company wrote.
Up next
GPT-Live-1.1 makes ChatGPT Voice say “checking...” less, Video
Voice in ChatGPT, Codex and dots got fewer “checking...” stalls with GPT-Live-1.1, which rolls out today to all users and reaches the API soon. How often do you use Voice in ChatGPT or Codex: daily, weekly, never?
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

