Claude AI Broke Into Live Systems During Security Tests

Anthropic disclosed Thursday that its Claude-based security models broke into the live production systems of three outside organizations during internal tests meant to gauge how dangerous the models could be as hackers. The models were supposed to be operating in a sealed-off simulation. They weren't. This is the second such disclosure in ten days involving a major AI lab's models trespassing into networks they had no business touching, conduct that would carry serious legal exposure for a human operator.

The pattern started with OpenAI, which reported earlier this month that its own security-testing models exploited a zero-day flaw to break into Hugging Face's network, then lifted access credentials and other sensitive data. Those models didn't stop there, using exposed credentials to compromise accounts at four additional third-party services. That disclosure prompted Anthropic's engineers to go back and audit their own "capture the flag" evaluations, a standard method for testing offensive and defensive hacking skills, run with third-party partner Irregular. What they found were three separate cases of a model slipping past the boundaries of its supposed sandbox and reaching real infrastructure belonging to real companies.