Anthropic gives METR employee-like access to evaluate its models

Anthropic said it had given six staff from the non-profit evaluator METR company hardware, office space, and access to internal tools and systems comparable to what Anthropic’s own risk-assessment teams use. The arrangement excludes customer data, attorney-client privileged material and model architecture details, but otherwise encourages METR staff to be informed as internal colleagues would be. METR can publish its own risk assessments, incident write-ups and practice analyses without Anthropic holding a veto; Anthropic may require redactions only from those three predefined categories, and METR is free to say publicly whether it thinks a given redaction was justified.

The step followed through on a commitment co-founder and chief executive Dario Amodei made in his September essay “We Must Pace the Frontier”, which proposed giving independent evaluators “employee-like access — desks, badges, laptops and the right to publish their findings” as a first, unilateral step toward making companies’ pacing claims externally checkable. Anthropic framed the move as part of its pledge, made alongside other signatories of the White House Accord on Superintelligence, to work with independent external auditors. “We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands,” the company said.

Anthropic named two other evaluators already under way: SecureBio, which is reviewing the chemical and biological weapons chapter of Anthropic’s August 2026 risk report after an earlier pilot review, and Accenture, paid in September to embed evaluators directly in model development, alongside the consultancy Faculty. Anthropic said it was in early talks with Hugging Face and others about similar arrangements.

The announcement arrived a day after METR itself published a framework proposing what access third-party assessors should have to judge loss-of-control risk at any developer, and which had stopped short of calling for literal employee status. Anthropic’s arrangement with METR goes further on hardware and systems access than that minimum ask, though it keeps the same carve-outs — architecture details and privileged material — that METR’s own proposal had flagged as short of true employee-level visibility.