Anthropic provides Model Evaluation and Threat Research with internal access to audit its AI models and safety practices

Meta description: Anthropic gives METR staff hardware, office space and internal tools to review AI models and safety practices, with independent findings publication.

Tags: Anthropic, METR, AI safety, model review, White House Accord

Featured image: Abstract technology style, layered review panels and access pathways over a neural network shape, no real people, no company logos, no text.

Anthropic has given six METR reviewers inside access to its AI models and safety practices, making the Anthropic METR inside access arrangement broader than earlier external reviews. The nonprofit’s staff receive company hardware, office space, and access to internal tools, a setup that Anthropic says puts them roughly on par with its own risk teams. At the same time, customer data, information covered by attorney-client privilege, and details about model architecture remain off-limits. The structure matters because it separates practical transparency from unrestricted disclosure: METR can review more closely, while Anthropic can request redactions in certain predefined categories.

Background: A commitment to outside review

The arrangement follows Anthropic’s signing of the White House Accord on Super Intelligence, which includes a commitment to working with outside reviewers. Anthropic had already engaged external parties, including SecureBio, Accenture, and Faculty. Those reviewers have so far received less access and tighter limits on what they can publish. That earlier pattern provides the baseline for the METR deal: it is not the first outside review effort, but it is described as going further than what the company has offered other outside reviewers.

Development: Six reviewers placed near Anthropic risk teams

Six METR staffers are the visible unit inside the arrangement. Their company-provided hardware, office space, and internal tool access are meant to let them work near Anthropic’s own risk teams rather than as external reviewers looking in from a distance. The access is still bounded. Customer data is excluded, as are materials covered by attorney-client privilege and details about the model architecture. Those exclusions define the review as a safety practice inspection, not a full public dump of Anthropic’s systems. The company’s position is that not all safety-relevant information can be made public. It argues that disclosure could help attackers or touch commercially sensitive material. The deal therefore gives reviewers a closer look while preserving categories Anthropic treats as sensitive.

Latest: Independent findings and redaction checks

METR will publish its findings independently, a central condition that distinguishes this arrangement from a purely internal safety report. Anthropic says it can ask for redactions in certain predefined categories, but METR is free to say publicly whether it thinks those redactions were justified. That creates a two-layer transparency model: the findings can be released, and the boundaries around release can themselves be reviewed. The latest expansion of access also sits alongside new external relationships. Anthropic has started early talks with Hugging Face.

What to watch next

The current stage is implementation, not final judgment. The company has announced the inside-access structure, and the next observable test is how METR’s independent publication process works when redaction requests appear. Readers can watch for three near-term markers: whether METR releases findings, how it describes any redactions it received, and whether the early talks with Hugging Face develop into another external review arrangement. For now, the factual record is limited to the access terms, the exclusions, and the commitment to independent publication.