Claude reached 3 real systems during cyber evals

Anthropic found three incidents in which Claude gained unauthorized access to three organizations’ real systems after reviewing 141,006 cybersecurity evaluation runs. Anthropic’s incident report attributes the failures to unintended internet access and says it stopped the evaluations, notified affected organizations, and is tightening monitoring and vendor controls. The lesson extends the model-sandbox containment problem: builders should treat evaluation harnesses for autonomous agents as production-grade security boundaries, not disposable test rigs.