Anthropic lets outside auditors check its work: what changed Oct 15
Anthropic gave a team from METR, a nonprofit AI evaluator, employee-like access to review its risk practices and publish findings Anthropic can't veto. Here's what that means for how much you can trust Claude's safety claims.
On October 15, 2026, Anthropic announced it has given six employees from METR — a nonprofit that evaluates frontier AI models — "employee-like access" to its internal tools and systems. METR can now publish its own analysis of Anthropic's risk levels, incidents, and practices. Anthropic cannot veto publication.
This matters for how you use Claude: when Anthropic makes a safety claim, you now have a way to check whether an outside party backs it up. This lesson is about where to find that check and what it actually covers.
The access has limits. Anthropic can still require METR to redact a small number of categories: information about customers, attorney-client privileged material, or model architecture details. But METR is free to say publicly that it disagrees with a redaction — it just can't reveal what was redacted.
Anthropic also named two other evaluators already in motion. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report, after already publishing a pilot review of the previous one. Accenture and Faculty are working toward a broader partnership that includes red-teaming models and testing safeguards. Anthropic described three access tiers for evaluators — narrow and approval-gated, broader with no approval needed, and METR's "employee-like" tier — and said it's in early talks with others, including Hugging Face.
Why this is useful to you: if you're citing Claude's safety posture to a boss, a client, or a compliance team, "Anthropic says so" is a weaker claim than "Anthropic says so, and METR — an independent nonprofit with no equity in Anthropic — can publish findings Anthropic can't block." Knowing the difference between a company's own claim and a verified claim is a skill you can ask Claude to help you apply, whether the topic is AI safety or a vendor's product spec.
What shipped: Anthropic gave METR employee-like access to review its risk practices, with publication rights Anthropic can't veto (narrow redaction categories only). SecureBio and Accenture/Faculty are earlier-stage evaluator partnerships. What to do with it: when a safety or capability claim matters for your decision, ask whether it's self-reported or independently checked — and name the checker.