Mythos Found Flaws in Sensitive U.S. Systems During Testing

Mythos identified vulnerabilities in sensitive U.S. government computer systems during a test with intelligence agencies, putting the model’s cybersecurity capabilities in focus as agencies assess how to use advanced AI tools.

Some flaws surfaced within hours. That did not mean the model could exploit the vulnerabilities in the same period.

The government tested Mythos through Project Glasswing, Anthropic’s defensive cybersecurity program. When Anthropic launched Glasswing in April, it offered select organizations Claude Mythos Preview for identifying and repairing security flaws in critical software. Mythos has identified thousands of severe vulnerabilities, and the preview is not slated for general release.

Anthropic later expanded the initiative to about 150 additional organizations across more than 15 countries. The participants include groups in power, water, health care, communications and hardware. Anthropic allocated as much as $100 million in model access credits and $4 million for open-source security groups.

At a June 11 hearing, Sen. Mark Warner, D-Va., said the tool “broke into almost all of our classified systems, not in weeks but in hours.” Warner attributed that information to Gen. Joshua Rudd, head of the National Security Agency (NSA) and U.S. Cyber Command.

Anthropic and the NSA declined to comment on the reported exercise.