Mythos AI cracks US classified systems in hours, raising AI alignment crisis

For AI researchers and safety advocates, the Mythos breakthrough is a stark warning. The model from Anthropic—a company founded on AI alignment—penetrated classified systems faster than human operators ever could, raising profound questions about whether safety methods can keep pace with capability. If well-intentioned AI can find flaws this quickly, adversarial models could wreak havoc.

On June 23, 2026, the Associated Press reported that Anthropic's latest AI model, Mythos, successfully identified vulnerabilities in highly sensitive classified U.S. government systems during a testing exercise, according to an anonymous U.S. official. The revelation underscores the dual-use potential of frontier AI systems: while capable of dramatically improving cybersecurity defense, the same technology poses immense risks if misused or if it falls into adversarial hands. The testing was part of Project Glasswing, an Anthropic-led initiative bringing together tech companies and government agencies to preemptively find and fix critical software vulnerabilities before they can be exploited by malicious actors. The official stressed that while the model identified vulnerabilities within hours, this did not necessarily mean it could exploit them in that timeframe, but the speed alone is alarming.

“

The NSA declined to comment, and Anthropic also refused to comment, reflecting the sensitive nature of the project and the escalating tensions between the AI startup and the Trump administration.

Senator Mark Warner, speaking at a June 11 Senate Banking Committee hearing, had already alluded to the results, claiming that “this tool broke into almost all of our classified systems, not in weeks but in hours,” attributing the information to General Joshua Rudd, head of the NSA and U.S. Cyber Command. The NSA declined to comment, and Anthropic also refused to comment, reflecting the sensitive nature of the project and the escalating tensions between the AI startup and the Trump administration.

Context is crucial: Anthropic, a San Francisco-based AI startup reportedly planning an IPO, has been locked in a dispute with the U.S. government over the military use of its models. The company refused to allow its AI to be used for domestic surveillance or fully autonomous weapons systems. In retaliation, the Trump administration placed Anthropic on a national security blacklist and earlier in June 2026 ordered it to suspend exports of its advanced models, including Mythos and Fable, to all foreign nations and foreign nationals. The New York Times further reported that the NSA lost access to Mythos amid the dispute, a significant setback for the intelligence community that had presumably been relying on the model for vulnerability detection.

The situation presents a classic dilemma for AI safety: Anthropic, founded with a mission to develop safe and aligned AI, refuses military applications that conflict with its principles, yet the government sees this as a risk to national security. The export restrictions could hamper Anthropic’s global business, potentially impacting its valuation and IPO prospects. Meanwhile, the fact that Mythos can so rapidly probe classified networks highlights the vulnerability of even the most secure systems to advanced AI. If a friendly model can do this in hours, a hostile state actor’s AI could potentially do worse.

The industry implications are broad. For AI startups, the episode demonstrates that cutting-edge capabilities will inevitably draw government attention, both collaborative and coercive. Navigating regulation—especially export controls and military contracting—will become a critical skill for AI founders. For the cybersecurity industry, AI-powered vulnerability discovery could revolutionize defensive operations, but it also means that patch timelines must shrink dramatically. The “hours not weeks” metric suggests a paradigm shift: traditional vulnerability management cycles are too slow.