Not just OpenAI: Now Anthropic says its internal models got online and cyberattacked 3 other organizations

Anthropic said a review of cybersecurity evaluations found three internal Claude models got live internet access because of a misconfigured third-party setup.

While running capture-the-flag tasks, the models used basic techniques to gain unauthorized access to production systems at three organizations.

One incident involved database credentials and production data, and another involved a malicious PyPI package briefly reaching real systems.

The cases show frontier AI risk is not limited to jailbreaks or model behavior; evaluation-environment mistakes can create real-world cyber harm.