AI · 17 October 2026 · 5 min read
Anthropic let six outside researchers into the building. It kept the right to decide what they must cut
METR staff now have hardware, desks and tools close to those of Anthropic’s own risk team. They can publish without a veto. Anthropic can still require redactions, and Anthropic decides what falls into the categories that must be cut.
Anthropic said on 15 October that it has given employee-like access to six people from METR, the nonprofit that evaluates frontier models. They have company hardware, office space, and internal tools roughly level with Anthropic’s own risk-assessment teams. Anthropic says that access stops short of customer data, material covered by attorney-client privilege, and information about its model architectures. METR is expected to publish its view of Anthropic’s risk levels, incidents and practices. Anthropic says it cannot veto that publication.
The sentence that limits the offer is the next one. Anthropic can require a redaction when the text falls into categories set in advance, including the ones above, and Anthropic decides whether a given passage is in those categories. METR keeps the right to say, in public, how much access it actually received and what it was not shown. A reader of the eventual report will have to take that statement seriously. A report that cannot mention the architecture, and whose author cannot overrule the company’s classification, is an outside report with a gate still held by the company.
No finding has been published
Nothing from METR under this arrangement is out yet. The announcement is the key, not a result. Anthropic also says it is working with SecureBio and Accenture as evaluators, and that it is in early talks with others, including Hugging Face. It describes three tiers. The narrowest needs Anthropic’s approval before publication. A middle tier can publish without approval, still subject to redaction. METR is the tier with the desks and the laptops.
This is the practical question left by the reported demand for safety records from the same company. An outside group that can sit with the incident log is more useful than a pledge to be audited. It becomes evidence when METR publishes a page that names an incident Anthropic did not, or says it was refused a system it asked for. Until that page exists, the news is an access agreement, with the redaction pen still in Anthropic’s hand.