Anthropic gives six METR evaluators employee-level access

Anthropic has given six researchers from METR, a nonprofit that tests advanced AI models, roughly the same access to its systems that its own risk teams have. It announced the step on October 15, 2026. Anthropic also gave up the right to block what METR publishes about its risks, incidents and safety practices. METR is the only group in the top tier of Anthropic's new system for outside evaluators.

What METR gets

"We've onboarded six METR employees with company hardware, office space, and access to internal tools and systems," Anthropic wrote. The access is roughly comparable to what its own risk assessment teams use, the company says.

METR will publish its own assessments. Anthropic cannot veto them, but it can require redactions in narrow categories. TechCrunch lists those categories as customer data, privileged material and details of model architecture. Anthropic's post adds that "METR is free to characterize our redactions publicly." In other words, readers will learn when something was held back, even if they cannot see it.

Three tiers of outside access

Anthropic now sorts its outside evaluators into three tiers, according to its post and TechCrunch. Each tier adds freedom or access:

Tier	Who is in it	What it allows
1	SecureBio, a biosecurity nonprofit	Publishing needs Anthropic's approval
2	Not named	Publishing without Anthropic's approval
3	METR, the only group so far	Free publishing plus employee-level access

SecureBio is reviewing the chapter on chemical and biological weapons in Anthropic's August Risk Report. Anthropic says it is also working with Accenture and its AI unit Faculty on a broader partnership.

Why Anthropic is doing this now

The move follows a run of incidents at AI labs. TechCrunch notes that labs have disclosed AI agent incidents "every week or two" since late August 2026. Anthropic, OpenAI, Google, Meta, Nvidia and xAI have also signed a White House accord on superintelligence. In it, the companies pledged to work with independent auditors or evaluators.

The idea of evaluators working inside the company is not new for Anthropic. In September, it named Accenture as its first embedded evaluator, The Standard reported, citing AFP. Anthropic and Accenture each expect to invest at least $1 billion over five years. At the time, Anthropic said it was in talks with METR about joining, using METR's own funding.

Funding is the sore point. METR says it accepts no money from AI companies or their executives, The Standard reports. More than 100 AI researchers, including Geoffrey Hinton, signed a letter asking that embedded evaluators be "meaningfully independent." METR itself is also among the groups named in a federal probe of AI agents that we covered on October 1.

"We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same," Anthropic wrote. TechCrunch reports that OpenAI has pledged a similar arrangement.

What this means for developers

For teams that build on Claude, nothing changes in the product today. The change is in what you will be able to read about the vendor.

METR's reports will be a new public source on how Anthropic handles incidents and risk. If your company runs vendor-risk or AI governance reviews, add them to the reading list once they appear. Pay attention to how often redactions are flagged and in which categories.

The tier system also gives buyers a simple question to ask any model provider. Who outside your company reviews your safety work, what can they see, and can they publish without your approval? Ask it of every vendor you rely on, not only Anthropic.

The next things to watch are METR's first published assessment and whether OpenAI's pledge turns into a matching arrangement. If both happen, outside review with publishing rights could become a normal line in enterprise AI contracts.