AI NewsInfrastructureAnnouncement
Anthropic gives METR employee access and no veto on publication
Anthropic has given six METR staff company laptops, office space and internal tool access to run their own risk assessments, and signed away its right to block publication of what they find.
Image: Anthropic
Why it mattersA buyer who ships model-backed features now has an outside risk report to point at when a lawyer or a customer asks who checked the vendor, instead of only the vendor's own safety page.
A frontier AI company giving an outside team the right to publish without approval is new. Anthropic said in a blog post that it has onboarded six METR employees with company hardware, office space and access to internal tools roughly on par with its own risk assessment teams, and that it cannot veto what METR publishes about Anthropic's risks, incidents or practices.
Anthropic's blog post calls METR a nonprofit that evaluates frontier AI models. The arrangement is the top of a three-tier system Anthropic is building for independent evaluators, and METR is the only team at that top tier so far.
What Anthropic can and cannot do
Anthropic can require METR to redact information that falls into narrow pre-defined categories: data about customers, attorney-client privileged material, and information about model architectures. METR is free to say in public that a redaction happened and whether it believes the redaction was justified, as long as it does not reveal the redacted content itself. METR's engagement runs under enforceable agreements, and other METR employees who work with Anthropic on separate projects are siloed from this team.
The three tiers
In the first tier, an evaluator runs a narrowly scoped review and needs Anthropic's approval to publish. SecureBio sits here, reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report after an earlier pilot. In the second tier, an evaluator can publish without approval and sees more information, with the same redaction categories available to Anthropic. The third tier adds the employee-like access and the publication rights that METR now has. Anthropic says it is also working toward partnerships with Accenture, Faculty and Hugging Face.
Why this shape
Anthropic uses a banking analogy: regulators place resident examiners inside the largest banks because the bank knows more about its own risk than any outsider can, and a bank's risk-taking can threaten the public. Giving an outside team deep access is the only way to close the gap between what Anthropic knows about its own systems and what an evaluator can learn by reading papers.
What matters here is that an AI company has written down, in public, a set of rules that limit its own power to block a finding. For a team buying model capacity for a feature that will carry risk, that matters. The vendor's own safety page is the vendor's own account of its safety; a report from an evaluator who cannot be blocked from publishing is a different kind of document. OpenAI promised the same arrangement on the day Dario Amodei first described the idea, and a White House accord on independent review has since been signed by Anthropic, OpenAI, Google, Meta, Nvidia and xAI, according to TechCrunch's report on the Anthropic announcement. The test is whether a report from one of these teams reaches the public with findings the vendor would rather not see, and whether the next buyer asks for one by name in a procurement checklist.
Source
Primary source: Anthropic. Reporting: TechCrunch.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
