An Anthropic AI model cracked into NSA classified systems during a controlled security drill - and the fallout is still being sorted out.
Sen. Mark Warner, citing a briefing from NSA and U.S. Cyber Command chief Gen. Joshua Rudd, told The Economist that Mythos breached "almost all" of the agency's classified systems in hours, not weeks. The evaluation happened on June 11, one day before the U.S. government issued a directive barring foreign nationals - including Anthropic's own non-citizen staff - from accessing Fable 5 and Mythos 5. Anthropic, unable to enforce a nationality-based access filter in practice, pulled the models globally. The Rudd quote, buried in a June 14 Economist report, went viral a week later before the article's author clarified that the "NSA got hacked" framing was inaccurate.
What the Rudd quote does supply is the context the government conspicuously withheld when it issued the ban - the first time the U.S. has applied export controls to an AI model rather than to the chips running it. Anthropic says the underlying issue was a narrow jailbreak, not autonomous offensive intrusion: the model was asked to analyze a codebase and flag vulnerabilities, and surfaced a handful of already-known bugs. The company also argues GPT-5.5 exhibits the same behavior, which is either a meaningful defense or a broader indictment, depending on how worried you are about the category.
Meanwhile, roughly six Anthropic engineers are reportedly embedded inside the NSA under Project Glasswing, adapting Mythos for operational use - work that may extend to probing networks in China and Iran. That context makes the public ban look less like a rupture and more like a negotiation in progress.