Anthropic is letting METR researchers look under the hood with employee-level access

TechCrunch reports Anthropic is giving six researchers from METR, a nonprofit that tests frontier models, employee-level access to independently assess models, safety practices, and risk management. The team gets company laptops, office space, and internal tools roughly on par with Anthropic’s own risk teams; staff are encouraged to share information as they would with a colleague on alignment risk. Anthropic likens it to resident bank examiners.

METR will publish its own assessments. Anthropic can’t veto publication but can require redactions in narrow categories — customer data, attorney-client privilege, model-architecture details — and METR may say publicly whether it believes redactions were justified. METR sits alone in the top tier of a three-tier evaluator system; lower tiers get narrower access and, at the bottom, need approval to publish.

This is one of the more serious transparency moves we’ve seen from a frontier lab — not a PDF scorecard, but people with badges and the right to publish.

It arrives after a stretch of agent-misbehavior disclosures roughly every week or two since late August, which sharpened the demand for outside eyes. Anthropic is also making good on Dario Amodei’s September essay proposing embedded evaluators; OpenAI pledged similarly, and a White House accord has major labs committing to independent review.

Caveats matter. Redaction categories that include model architecture can hide the guts of how risk is measured. Six researchers is a start, not a regulator. Still, giving up veto power is the part that changes incentives. I’m watching METR’s first public assessments under this arrangement, and whether OpenAI and Google match the employee-level tier rather than the press-release tier.

SecureBio is cited as an example of a bottom-tier evaluator running narrow reviews that require Anthropic approval to publish. The middle tier gets more information and can publish without approval.