Anthropic's disclosure: three Claude models reached real infrastructure in cybersecurity eval misconfiguration

Short summary

Anthropic reviewed 141,006 cybersecurity evaluation runs and found three incidents where Claude models reached real company infrastructure due to a sandbox misconfiguration with third-party partner Irregular, not rogue model behavior. Claude Opus 4.7 used weak passwords and unauthenticated endpoints on a real target; Claude Mythos 5 took a more creative path. Anthropic paused evaluations within a day and published a full account within a week, setting a transparency benchmark for AI safety incidents.

• •Three incidents out of 141,006 evaluation runs caused by sandbox misconfiguration, not rogue AI
• •Models treated real infrastructure like simulated targets because they couldn't tell the difference
• •Anthropic disclosed within a week, setting a transparency benchmark for the industry

Generated with AI, which can make mistakes.