Anthropic's frontier AI model, Mythos, has demonstrated a startling capability: it identified vulnerabilities in highly classified U.S. government computer systems within hours during a secret testing program called Project Glasswing, the Associated Press reported on June 23, 2026.
The revelation came from NSA Chief Joshua Rudd, who told a congressional hearing that Mythos "broke into almost all of our classified systems, not in weeks, but in hours."
The exercise — conducted in partnership between Anthropic and U.S. intelligence agencies — was designed to find and fix critical software vulnerabilities before malicious actors could exploit them. Senator Mark Warner of Virginia referenced the testing in a congressional hearing, saying he had been briefed by Rudd on the remarkable results.
However, the breakthrough arrives amid a deeply strained relationship between the AI company and Washington. The U.S. government recently placed Anthropic on a national security blacklist after the company refused to allow its models to be used for domestic surveillance and fully autonomous weapons systems. The government has also ordered Anthropic to suspend global exports of its Mythos and Fable models to destinations worldwide and to all foreign nationals, citing national security concerns.
Anthropic's relationship with the Pentagon had already been rocky. The company stated it was willing to adapt its policies but drew a firm line at mass surveillance and autonomous weapons. The NSA reportedly lost access to Mythos amid the dispute, as reported by the New York Times.
The episode highlights a defining tension of the age: as AI systems become extraordinarily powerful tools for both offensive and defensive cybersecurity, governments must decide whether to embrace, restrict, or fear the very technology they helped incubate. The question of who gets to control — and contain — frontier AI capabilities is no longer theoretical. It is playing out in real time between Silicon Valley and the intelligence community.
In a follow-up development, the U.S. has since allowed Anthropic to release Mythos to 'trusted' U.S. organizations, suggesting a calibrated approach to access rather than outright prohibition. But the broader debate over AI safety, export controls, and national security is far from settled.
