From Anthropic · · 50s
Anthropic's independent evaluators update
Anthropic starts giving employee-like access to METR for AI safety assessments.
In brief
Anthropic starts giving employee-like access to METR for AI safety assessments. Anthropic began embedding METR evaluators with employee-like access to assess AI risks while maintaining data security. Originally reported by Anthropic.
Announcements
Anthropic announced employee-like access for METR evaluators to assess AI safety risks.
Why independent evaluators
Anthropic needs third-party risk assessments for powerful AI models without sharing sensitive data.
METR access details
Six METR employees get company hardware, office space, and internal tools with limited data sharing rules.
Publication rights
METR publishes risk assessments but can redact sensitive information per pre-defined categories.
Other evaluator partnerships
Anthropic works with SecureBio and Accenture for limited AI safety evaluations.
Three-tier access system
Anthropic develops three tiers of evaluator access with increasing publication rights and data access.
The key point
Anthropic began embedding METR evaluators with employee-like access to assess AI risks while maintaining data security.
