Mythos AI turned hours into zero-day discovery for US classified systems

For cybersecurity leaders, the speed at which Mythos dissected classified systems is a harbinger of the AI-powered vulnerability hunting era. Traditional penetration testing would take weeks to map even a fraction of what the model uncovered in hours. This test validates that AI can now operate as a near-instantaneous, scalable red team, forcing organizations to rethink vulnerability management lifecycles and defensive automation.

In a revelation that underscores the accelerating cybersecurity threat landscape, a U.S. official confirmed that Anthropic’s advanced AI model, Mythos, successfully identified vulnerabilities in highly classified U.S. government computer systems during a joint testing exercise with intelligence agencies. The exercise, part of an initiative called Project Glasswing—a collaborative effort involving tech giants to secure critical software against severe AI-driven fallout—demonstrated that the model could pinpoint weaknesses not over weeks but within hours. Senator Mark Warner (D-VA) disclosed the findings during a June 11 Senate Banking Committee hearing, citing information from General Joshua Rudd, head of the NSA and U.S. Cyber Command. 'This tool broke into almost all of our classified systems, not in weeks but in hours,' Warner stated, though officials later clarified that while the model identified vulnerabilities, it did not exploit them in that timeframe.

“

Conversely, the administration early in June issued a directive requiring Anthropic to bar foreign nationals from using its latest models, Fable 5 and Mythos 5—a move that hints at fears of adversarial access and industrial espionage.

The test’s implications are profound. It signals a paradigm shift in offensive and defensive cyber capabilities: an AI system, likely trained on vast codebases and attack patterns, can systematically probe hardened, air-gapped networks at machine speed. The fact that the model was able to find flaws across 'almost all' tested systems suggests systemic weaknesses that human red teams might take weeks or months to uncover. This speed advantage could compress vulnerability discovery lifecycles from months to days, fundamentally altering the time pressure on security operations centers and vulnerability disclosure processes.

Contextually, the test occurred amid a delicate dance between Anthropic and the Trump administration. While the company cooperated with intelligence agencies, broader tensions simmered over how its AI would be used in military contexts. Anthropic has publicly expressed concerns about the militarization of AI, pushing back against unrestricted military applications. Conversely, the administration early in June issued a directive requiring Anthropic to bar foreign nationals from using its latest models, Fable 5 and Mythos 5—a move that hints at fears of adversarial access and industrial espionage. This friction illustrates the classic national security dilemma: the very tools that can strengthen defense are also the ones that, if misused or exported, could erode security.

The testing exercise was framed as a preemptive security measure under Project Glasswing, which aims to harden critical software before bad actors can leverage AI for catastrophic attacks. However, the revelation that an AI model found widespread vulnerabilities in systems protected by the highest classifications raises uncomfortable questions: how many of these flaws have existed unnoticed? Were any already exploited by state-sponsored attackers? And what does Anthropic’s discovery mean for the security posture of allies who may not have access to such advanced testing?