Anthropic gives METR employee-level access to evaluate its systems' safety

On October 15, 2026, Anthropic announced that it will give a team from METR employee-level access to evaluate the safety of its systems. METR is a nonprofit that specializes in testing frontier models. The company is also in early talks with several other organizations, and Hugging Face is the one it named. This is the latest of 51 entries in the Policy section of Anthropic's website.

What exactly was announced

The phrase "employee-like access" matters. External evaluators usually work through an API or closed test environments, where they see only what the lab has chosen to show. Access close to that of an employee implies a deeper look into the systems. The announcement does not say which models, tools, and internal data will fall inside this environment. It also gives no timelines and does not describe how results will be published.

The list of partners currently includes one specific participant, METR. All that is known about the others is that talks are at an early stage. Hugging Face is named explicitly; the other names have not been disclosed.

• Announcement date: October 15, 2026
• Partner with access: the METR team, a nonprofit evaluator of frontier models
• Format: access comparable to that of employees
• Candidates for expansion: several organizations, including Hugging Face
• Status of other agreements: early talks
• Publication of results and list of models: not disclosed

Context: an autumn packed with policy

The announcement sits alongside other Anthropic moves in 2026. An updated Usage Policy came out on October 8. In February, the company introduced Responsible Scaling Policy 3.0, a document on reducing catastrophic AI risks. In December 2025, it published the Frontier Compliance Framework for California's Transparency in Frontier AI Act, and in September 2025 it backed the SB 53 bill.

On top of that, the company is engaging with government and voters. Anthropic put $20 million into Public First Action in February and another $20 million in July, for a total of $40 million. On June 12, the company issued a statement on a US government directive to suspend access to Fable 5 and Mythos 5 under export controls. In February and March, Dario Amodei made statements on negotiations with the Department of War. Separate pieces cover data centers: support for Governor Abbott's calls for responsible construction, and a commitment to cover increases in consumer electricity rates caused by the company's own data centers.

Against this backdrop, independent evaluation looks like a logical addition. When a company publishes framework documents and agrees to regulatory requirements, external review stops being a goodwill gesture and becomes a working tool for confirming what the company claims.

What the announcement leaves out

The main question is about boundaries. Can evaluators publish their findings without Anthropic's sign-off? Do they have the right to halt a release, or at least to object publicly? Do they get access to unannounced models before release? The announcement answers none of these questions, and the answers determine how independent the review is in practice.

The second question concerns who takes part. Including Hugging Face, a platform with a strong open-source community, could broaden the range of evaluation methods. For now, though, these are only discussions.

Practical takeaway

For teams that buy or embed Anthropic models, the signal so far is more about process than technology. Watch how METR and other evaluators publish their results: their reports will become an additional input for internal risk assessments. Compliance staff and lawyers working under California requirements should check the wording of the new October 8 Usage Policy and RSP 3.0 against what the evaluators confirm independently. For everyone else, it is enough to note this: Anthropic now has an external review channel with deep access, but its real value will depend on terms the company has not yet disclosed.