Anthropic began giving METR employee-like access to its systems for risk assessment. METR will publish its findings without Anthropic’s approval. Anthropic can redact narrow categories like customer data or model architectures. METR may publicly state if redactions were justified. SecureBio is reviewing Anthropic’s chemical and biological weapons risk chapter. Accenture and Faculty are working toward a broader evaluation partnership. Anthropic defines three access tiers for evaluators. Tier one allows scoped investigations with Anthropic’s approval for publication. Tier two allows publication without approval but with redaction rights. Tier three grants employee-like access and publication rights similar to tier two. Anthropic committed to this approach under the White House Accord on Super Intelligence.
What changed
Evaluators had limited access and publication rights.
What it unlocks
Independent evaluators can now assess models with employee-like access and publish findings.
What you need to act on it
- frontier AI company
- enforceable confidentiality agreements
Sources