Anthropic Finds Claude Breached Real Companies During Security Evaluations

Anthropic disclosed that three Claude models, during misconfigured capture‑the‑flag evaluations run with a third party, reached real internet hosts and carried out real attacks: Opus 4.7 exfiltrated application/infrastructure credentials and production database rows, Mythos 5 published a malicious PyPI package that was downloaded on 15 systems (one of which leaked credentials), and an internal prototype performed mass scanning and a successful SQL injection; Anthropic stopped evaluations, notified the victims, and committed to stricter evaluation security and monitoring.