Analysis
Anthropic is giving six researchers from the nonprofit AI-safety group METR employee-level access to its operations, company laptops, office space and internal tools on par with Anthropic's own risk-assessment teams, so METR can independently evaluate its models and safety practices, TechCrunch reported Thursday. METR will publish its own assessments, and Anthropic has given up the right to block that publication except for narrow, pre-agreed redaction categories, a materially stronger commitment than the access-and-NDA arrangements most AI labs offer outside evaluators.
A test Anthropic set up for itself
The move follows Anthropic CEO Dario Amodei's September essay pledging deeper third-party scrutiny, and sits downstream of the White House AI safety accord signed earlier this year by Anthropic, OpenAI, Google, Meta, Nvidia and xAI, a voluntary framework with no enforcement mechanism of its own. Giving METR this level of access is Anthropic's attempt to make that pledge concrete rather than rhetorical, and it puts informal pressure on the other accord signatories to offer something comparable or explain why they haven't.
METR isn't a soft choice of auditor. Pulse has covered the group's prior work flagging AI agent-control problems and benchmark-gaming behavior in frontier models, and, notably, a security incident in which a stolen Anthropic API key was used to burn through roughly $600,000 in credits on METR's own infrastructure. That history cuts both ways: it shows METR is willing to publish uncomfortable findings about the labs it evaluates, but it also means this new access arrangement has to directly address how Anthropic secures its relationship with an evaluator that's already been a target once.
Anthropic frames the move as accountability: "We want to be publicly accountable for the decisions we're making," the company said in confirming the arrangement. For VCs and enterprise buyers evaluating frontier-model vendors, the practical test isn't the access grant, plenty of companies announce evaluator relationships, it's whether METR's published findings over the next few quarters actually surface something Anthropic would rather have kept quiet. If they do, this becomes a genuine differentiator against OpenAI and Google, which have less formalized third-party review processes. If METR's reports stay anodyne, the access grant reads more as a well-timed move landing right as the FTC is already demanding safety-incident documents from Anthropic and OpenAI. It's also self-selected oversight: Anthropic chose METR and is the one granting access, a different thing from a regulator or an independent board compelling the review.
