Anthropic has embedded six evaluators from the nonprofit METR inside the company with employee-like access, and METR can publish what it finds without asking Anthropic first. Why it matters: independent safety checks usually happen from the outside, with the lab controlling what reviewers see. This arrangement moves the reviewer inside the building.
October 15, 2026 — Anthropic said on Thursday that a small team from METR, a nonprofit that evaluates frontier AI models, has been onboarded with company hardware, office space, and access to internal tools and systems.
What happened?
According to Anthropic, the six-person METR team now has access broadly comparable to what the company’s own risk assessment teams work with. The evaluators are there to independently assess advanced models, safety practices, and risk management — not just a finished model handed over for testing, but the processes around it.
The move follows a commitment Anthropic made in September to stronger independent, third-party assessment, and it follows the company’s signing of the White House accord in which AI companies pledged to work with independent external auditors or evaluators. Anthropic said it has also worked with METR over the past year on reviews of its Risk Reports.
What are the publication rules?
The part that makes this different from a normal consulting review is the publishing right. METR can release independent assessments of risks, incidents, and safety practices without prior approval from Anthropic. The company can require only limited redactions, covering areas such as customer data, privileged information, and certain details about model architecture.
Anthropic is also building out a tiered system for other evaluators. Some reviewers will work on narrowly scoped investigations with more limited access and publication rights, while the company says it hopes over time to have several evaluator teams at the deepest, employee-like level, each publishing its own analysis. SecureBio is already reviewing part of a recent Anthropic risk report at a lower tier, and Anthropic said it is working toward a broader evaluation partnership involving Accenture’s Faculty business.
Why does independent access matter?
Researchers and evaluators have raised a practical concern in recent months: if a lab decides what an outside reviewer can see, and whether the review can be published, the review is only partly independent. Giving an embedded team internal access, and a right to publish, is an attempt to answer that concern in a verifiable way rather than with a promise.
There are limits worth keeping in mind. The team is small — six people — the access is granted by Anthropic itself, and some categories of information stay off limits. Whether this becomes a real check will depend on what METR actually publishes, and whether other frontier labs match the arrangement, as Anthropic is urging them to do.
Why it matters
For readers, the takeaway is simple: safety claims from AI labs are about to get a new kind of test. If embedded evaluators start publishing findings that labs did not approve in advance — including uncomfortable ones — that gives customers, regulators, and investors something firmer to judge than a company’s own report about itself. That is the “everything AI, tested” standard applied inside the lab, not just to the product. Watch for METR’s first published assessment under this arrangement: its tone will tell you more than this announcement does.
Frequently asked questions
Who is METR?
METR is a nonprofit organization that evaluates frontier AI models, focusing on advanced capabilities, risks, incidents, and safety practices.
What access did Anthropic give METR?
Six METR employees received company hardware, office space, and access to internal tools and systems broadly comparable to Anthropic’s own risk assessment teams, with limited exceptions such as customer data and privileged information.
Can Anthropic block METR from publishing?
No prior approval is required for METR to publish its assessments. Anthropic can require only limited redactions in defined categories, such as customer data, privileged information, and certain model architecture details.
Sources: Reuters, Unite.AI