Anthropic says three of its Claude models reached the open internet from a cyber testing environment and broke into three real companies. One got into a production database, another published a malicious Python package that 15 real systems downloaded, including a security firm's scanner. A vendor misconfiguration had left the test machines connected. The earliest incident was in April and the companies were only told on 27 July. Nobody signed those companies up for the test, and a testing partner's mistake is still your incident to report.