Anthropic gives outside safety researchers the same access as its own staff

Anthropic said Thursday it is giving six researchers from METR, a nonprofit that tests frontier AI models, the same internal access as its own employees. The researchers get company-issued laptops, office space, and entry to internal tools roughly on par with what Anthropic’s own risk assessment teams use, according to a blog post from the company.

The move comes as reports of AI agents acting outside their instructions have surfaced every week or two since late August, pressure that has sharpened calls for independent checks on AI companies. Anthropic said employees are encouraged to share information with the METR team “just as they would with a fellow employee working on alignment risk,” a reference to checking that a model’s goals match the behavior it was trained for.

Anthropic also gave up its right to block what METR publishes. The nonprofit will issue its own assessments of Anthropic’s risks, incidents, and practices. Anthropic can still require redactions, but only in narrow categories it defined in advance, such as customer data, attorney-client privileged material, or details about model architecture. METR can publicly say whether it thinks a redaction was justified, as long as it doesn’t reveal what was redacted.

Anthropic compared the arrangement to bank oversight, pointing to how regulators place resident examiners inside the largest banks to monitor them from within.

METR is the first and only evaluator to reach the top of a three-tier review system Anthropic is building. Evaluators in the lowest tier, among them biosecurity nonprofit SecureBio, run narrower reviews and need Anthropic’s sign-off before publishing. A middle tier gets broader access and can publish without approval. METR is alone at the top tier, with staff-level access and no veto power over its findings.

The arrangement fulfills a pledge Anthropic CEO Dario Amodei made in a September essay, where he floated embedding outside evaluators inside AI labs as an early step toward slowing the pace of the technology’s development. OpenAI said it would adopt a similar approach shortly after. Anthropic, OpenAI, Google, Meta, Nvidia, and xAI have since signed a White House accord committing to independent review of their systems.

What remains unclear is how the other tiers will fill out, and whether other evaluators will get the same employee-level access METR now has. Anthropic’s blog post did not name other candidates for the top tier, and it did not say when METR’s first public assessment might arrive.