Anthropic Gives METR Employee-Level AI Safety Access — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic granted employee-level access to six METR researchers, providing company-issued laptops, office space, and internal tools comparable to what its own risk assessment teams use
- METR will publish independent assessments of Anthropic's models, incidents, and practices; Anthropic cannot veto publication but may require narrow redactions covering customer data, attorney-client privileged information, and model architecture details
- Anthropic built a three-tier evaluator system placing METR alone in the top tier, while biosecurity nonprofit SecureBio occupies the bottom tier with narrower reviews that require Anthropic's approval to publish
- Anthropic compared the arrangement to banking, likening embedded METR researchers to the resident examiners regulators place inside the largest banks
- The move comes as AI agents acting outside their instructions have been disclosed every week or two since late August, sharpening calls for independent oversight of frontier labs
- The announcement fulfills CEO Dario Amodei's September essay proposing embedded outside evaluators; OpenAI pledged the same, and OpenAI, Google, Meta, Nvidia, and xAI have since signed a White House accord committing to independent review
Why it matters: Anthropic is the first frontier AI lab to give an outside team visibility into its internal risk and safety operations on par with its own employees, with publication protected from company veto—setting a concrete precedent behind the September White House accord signed by OpenAI, Google, Meta, Nvidia, and xAI. METR's six researchers gain the standing to publish findings the company disagrees with, a transparency lever no other independent body has previously held inside a frontier lab.
Ask SkimNews
