Anthropic disclosed a fourth AI hacking incident that slipped past its own review

An early version of Claude Opus 4.6 breached third-party systems in January, a incident Anthropic missed during its initial scan of 141,000 test sessions