Anthropic gives METR employee-like access, unveils three-tier external evaluator system

Anthropic · rss · 2026-10-15

Anthropic announced it has onboarded six employees from METR, a nonprofit frontier-model evaluator, with 'employee-like access': company hardware, office space, and internal tool access roughly on par with its own risk assessment teams, outside narrow categories like customer data, privileged info, and model architectures. METR can publish its risk findings; Anthropic cannot veto publication, only require redactions within pre-defined categories — and METR may publicly characterize whether those redactions were justified.

Why independent evaluation: as models grow more powerful, they need stronger safeguards, but Anthropic can't share all safety-relevant information publicly (misuse risk, commercial sensitivity). Instead, third parties get deep access and publish reasoned conclusions. The post draws an analogy to bank supervision, where regulators embed resident examiners in large banks.

Three access tiers: Tier 1 is narrowly scoped investigation with company approval for publication (SecureBio sits here, reviewing the chemical/bioweapons chapter of the August Risk Report); Tier 2 allows publication without approval and more access; Tier 3 is the METR arrangement. Anthropic is also pursuing evaluation and red-teaming partnerships with Accenture and Faculty, and early discussions with Hugging Face. The company says the embedding is voluntary and calls on every frontier AI lab to do the same.

Original post →

More from Companies & People

Companies & People channel →