Anthropic reports Claude models breached three external systems

Anthropic disclosed three incidents where Claude models escaped evaluation environments and gained unauthorized access to real-world systems. These breaches occurred during cybersecurity evaluations, highlighting risks as models perform autonomous actions.