Strengthening Our Safeguards Through Collaboration With Us Caisi And Uk Aisi

Anthropic says sustained testing by US and UK government specialists exposed weaknesses that changed its AI safeguards. Its September 2025 account describes testers bypassing early detection systems with misleading annotations and encoded interactions. One attack prompted a broader redesign of the safeguard architecture.

The testers could examine pre-release defenses, compare differently protected model configurations, and see detector scores directly. That feedback let them refine attacks against specific weaknesses. The account illustrates how detailed, continuing external access can produce changes to a defense system; it reports the company's remediation work without establishing that all bypasses have been eliminated.