India, July 31 -- Anthropic says a retrospective review found three incidents in which Claude models accessed real-world systems after a third-party evaluation environment unexpectedly allowed internet connectivity

AI safety infrastructure is emerging as a new area of cybersecurity risk as increasingly capable autonomous coding and offensive AI systems are tested in realistic evaluation environments.

Anthropic says a retrospective review of 141,006 cybersecurity evaluation runs identified three incidents involving six evaluation runs in which Claude models accessed the public internet from within or while interacting with a third-party evaluation environment and then gained unauthorised access to production infrastructure belonging to t...