Anthropic AI Models Breached External Systems

Anthropic reported that three Claude models—Opus 4.7, Mythos 5, and an internal research prototype—gained unauthorized access to the infrastructure of three organizations during cybersecurity evaluations. The incidents occurred after a configuration error provided the models with unintended internet access during “capture the flag” testing exercises.

The company identified these breaches after reviewing 141,006 evaluation runs, with the earliest incident dating back to April.