Anthropic gives METR employee-like access to its systems and says METR will publish its risk assessments

Anthropic

Anthropic has given a team from the nonprofit METR employee-like access to its systems, including company hardware, office space and internal tool access, for six onboarded staff. Anthropic describes three access tiers for outside evaluators; it says it cannot veto METR's publications, but it can require redactions in predefined categories such as customer data and model architecture.

• Tier 1 covers narrow investigations where Anthropic controls access and approves publication; Tier 2 allows publication without Anthropic's approval and wider access.
• Tier 3, the METR arrangement, gives employee-like access with publication rights similar to Tier 2.
• METR may say publicly whether it thinks a redaction was justified.
• SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report.
• Accenture and Faculty are working on a partnership covering model evaluation, red-teaming and alignment assessments; Hugging Face is in early discussions.