One of Anthropic’s AI models, Mythos, spotted weak spots in top‐secret U.S. government systems in a matter of hours, a U.S. official told The Associated Press. Working under Project Glasswing—a partnership between Anthropic and U.S. intelligence agencies—the model flagged vulnerabilities but didn’t actually exploit them. The effort aims to shore up critical software before any flaws can be weaponized against public safety, national security or the economy. Sen. Mark Warner of Virginia mentioned the exercise during a June 11 Senate hearing, quoting NSA and Cyber Command head Gen. Joshua Rudd: “This tool broke into almost all of our classified systems, not in weeks but in hours.” Neither the NSA nor Anthropic responded to requests for comment. The test results underscore the potential power of advanced AI to both find and, if misused, exploit software gaps at machine speed. Tensions have grown between Anthropic and the Trump administration amid fears over military and foreign use of its models. Earlier this month, the administration ordered Anthropic to block access for foreign nationals to its latest releases, Mythos 5 and Fable 5. Anthropic complied, disabling the models for all customers, arguing the restriction went beyond any real security threat it raised. Over 100 cybersecurity leaders—including executives from Adobe and Nvidia—urge the administration to reverse the ban. In a letter they say Mythos is “quite good” at uncovering and weaponizing software flaws but not uniquely so; other foundation and open‐source models perform similar security auditing tasks. They warn that sidelining these advanced tools could weaken U.S. defenses just as rival states ramp up their own AI capabilities.