Mythos AI Breached Classified Systems in Hours, Raising Space Defense Alarms

For the space defense community, the Mythos breach is a klaxon. Classified U.S. systems are the backbone of satellite operations, strategic missile defense, and space situational awareness—and an AI model capable of pinpointing vulnerabilities in mere hours shatters traditional defense-in-depth timelines. If a controlled test could achieve that, adversarial AI tools might already be probing the same systems, making the protection of orbital assets and ground stations an unprecedented urgent priority.

Anthropic’s Mythos AI model has demonstrated the ability to identify vulnerabilities within highly sensitive and classified U.S. government computer systems during a collaborative testing exercise, a U.S. official confirmed to the Associated Press on June 24, 2026. The revelation, first hinted at by Senator Mark Warner during a June 11 Senate Banking Committee hearing, underscores the profound double-edged nature of advanced AI in national security: the same models that promise to accelerate defensive cybersecurity can serve as potent offensive tools, raising urgent policy and ethical questions. The testing was conducted under Anthropic’s Project Glasswing, an initiative that brings together technology firms and government partners to assess and mitigate the systemic risks posed by frontier AI models to critical infrastructure and public safety. According to the anonymous official, Mythos identified certain weaknesses within hours, though this did not equate to successful exploitation. Senator Warner’s more dramatic characterization—that the tool “broke into almost all of our classified systems, not in weeks but in hours”—amplifies the perceived severity, and he attributed this assessment to General Joshua Rudd, head of the National Security Agency and U.S. Cyber Command. The NSA has declined comment, and Anthropic has remained silent, leaving the technical details and the extent of access unverified.

“

Anthropic’s Mythos AI model has demonstrated the ability to identify vulnerabilities within highly sensitive and classified U.S.

The story arrives amid escalating tensions between the California-based AI lab and the Trump administration. Only weeks before the disclosure, the administration issued a directive restricting the use of certain Anthropic models, while the company has voiced objections to the military deployment of its technology. This friction complicates the natural partnership between government and leading AI developers for defensive purposes. Project Glasswing itself symbolizes a proactive posture—enlisting top-tier AI to stress-test national security networks—but the results inevitably raise the specter of adversarial states catching up. If a U.S.-controlled model can sniff out vulnerabilities this efficiently, foreign adversaries or malicious non-state actors working with similar capabilities could pose an existential threat to classified infrastructure, including nuclear command and control, intelligence databases, and satellite networks.