Anthropic gives 6 METR evaluators internal access

What happened

Anthropic published “An update on independent evaluators” on October 15, 2026, at 17:20 UTC. The text was modified that same night, at 23:35 UTC, to clarify the terms of a SecureBio project.
A team of 6 METR employees, a nonprofit that evaluates frontier models, received access similar to an employee’s: company teams, an office and internal tools comparable to those of its own risk-evaluation teams.
METR will publish its findings. Anthropic cannot veto them, but it can require redactions in predefined categories, such as user data, privileged legal information and details of the model’s architecture. METR can say publicly whether it considers each redaction justified.
SecureBio is reviewing the chemical and biological weapons chapter of the August 2026 risk report. Accenture, with Faculty, is seeking a broader partnership: model evaluation, adversarial testing, alignment evaluations and safeguards testing. Initial conversations are under way with other organizations, among them Hugging Face.

The company describes three tiers of access. In the first, narrowly scoped research, Anthropic controls access and approves publication. In the other two, publication requires no approval, although it allows redactions. Anyone reading an evaluator’s report should know which tier it was done under.
Among the redactable categories is the model’s architecture. That is exactly what a third party would need to explain why a model fails. The safeguard that METR can comment on each redaction is what makes the limit verifiable.
For a Chilean company that buys Anthropic models, the change is one of evidence: there will be third-party reports with real access that can be cited in a vendor evaluation, instead of relying only on the manufacturer’s internal tests. The independence criteria (no equity stake and no financial or operational dependence) are declared by the company itself.

6: the METR employees with access similar to that of an Anthropic employee.

The publication cites Anthropic’s commitment, the previous month, to strengthen independent evaluation, and its adherence to the White House Superintelligence Agreement. OpenAI has also staked out its position: OpenAI publishes principles for external audits. On the failures the company has already documented, see Anthropic detailed 4 cases of improper access by Claude.

METR’s and SecureBio’s reports, with no publication date.
The partnership with Accenture and Faculty, under negotiation, with no timeline.
Conversations with other organizations, among them Hugging Face, with no timeline.

Anthropic argues that every frontier company should do the same. So far, OpenAI has published principles and Anthropic has published access.

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.