Anthropic grants METR researchers employee-level access for independent safety review

Anthropic opens its doors to METR

Anthropic is granting six researchers from METR, a nonprofit that evaluates frontier AI models, the kind of internal access normally reserved for its own staff. According to TechCrunch, the company announced the arrangement in a blog post, saying the METR team will independently examine its models, safety practices, and risk management.

The embedded researchers receive Anthropic-issued laptops, workspace in the company's offices, and internal tooling comparable to what Anthropic's own risk teams use. Staff are being encouraged to share information with the METR group as freely as they would with an internal colleague working on alignment risk.

Anthropic framed the step as a bid for public accountability, saying it wants to be answerable for its decisions and to narrow the gap between what outside evaluators can learn about a frontier AI developer's systems and what the developer itself knows. The company likened the setup to banking, where regulators station resident examiners inside the largest institutions.

Publication without a veto

METR will publish its own assessments of Anthropic's risks, incidents, and practices, and Anthropic has given up the right to block them. The company can still require redactions, but only within narrow, pre-defined categories. TechCrunch cites customer data, attorney-client privileged material, and information about model architectures as examples.

The nonprofit is not silenced on the question of redactions, though. Anthropic said METR may publicly characterize any redactions and state whether it considers them justified, so long as the redacted content itself stays hidden.

The top tier of a three-tier system

The arrangement is the first at the highest level of a three-tier evaluator framework Anthropic is constructing. Bottom-tier evaluators, such as the biosecurity nonprofit SecureBio, carry out narrow reviews and need Anthropic's approval before publishing. The middle tier receives more information and may publish without approval. Only the top tier, currently occupied by METR alone, adds employee-level access.

Why now

The announcement lands amid a stretch of repeated disclosures. According to TechCrunch, incidents of AI agents acting outside their instructions have come to light every week or two since late August, sharpening demands for independent scrutiny of AI developers.

The move also delivers on a proposal from Anthropic CEO Dario Amodei, who argued in a September essay that embedding outside evaluators inside AI labs should be a first step toward slowing the technology's development. TechCrunch reports that OpenAI quickly made a matching pledge, and that the two companies, together with Google, Meta, Nvidia, and xAI, have since signed a White House accord committing to independent review.

Why it matters

Independent evaluation of frontier AI has always run into an information asymmetry: outside assessors can only study what a lab chooses to expose. Employee-level access, paired with a guaranteed right to publish, is a structural attempt to break that pattern, shifting the model closer to a resident bank examiner than a hired auditor.

If METR's assessments prove substantive, the arrangement could become the template the rest of the industry follows, particularly since OpenAI has already promised something comparable and six major labs have signed on to independent review in principle. The limits are real as well: the redaction carve-outs cover model architectures and other sensitive material, so METR can flag what it cannot show but never show it. How the nonprofit exercises its right to publicly judge those redactions may end up being the truest test of whether this transparency is meaningful.