Did Anthropic’s Mythos Just Scare Washington? The NSA Breach Claim That Turned AI Into a National-Security Crisis
A reported claim that Mythos broke into classified NSA systems has transformed the Anthropic shutdown story from a tech dispute into a geopolitical warning about frontier AI.
The Anthropic controversy is no longer just a story about one company, one model, or one angry White House order. It is becoming a test case for the most uncomfortable question in technology: what happens when the same AI tools governments want to use for cyber defense are also powerful enough to threaten the government itself?
The most explosive claim now circulating comes from reporting around The Economist and subsequent discussion by national-security observers. Senator Mark Warner, the vice chair of the Senate Intelligence Committee, was quoted as saying that General Joshua Rudd, who leads the NSA and U.S. Cyber Command, told him that Anthropic’s Mythos had broken into “almost all” classified systems within hours. If accurate, that statement would mark one of the most dramatic public acknowledgments yet that frontier AI is crossing from software assistance into offensive cyber capability.
But the story is still contested. One version says Washington acted because the government saw an immediate national-security threat: a model that could exploit systems at speed, scale and depth beyond ordinary human teams. In that interpretation, Trump’s order restricting foreign access to Anthropic’s Mythos and Fable models was not a political tantrum. It was an emergency move after a government red line had already been crossed.
The other version is much less cinematic. Anthropic has reportedly argued that the problem involved jailbreaks and misuse risks that are not unique to its systems. The company’s defenders say the shutdown was disproportionate, opaque and possibly driven by panic inside a government that wants access to powerful AI but does not yet know how to regulate it. Under this view, the danger is not simply Mythos. The danger is an improvised licensing regime created through executive pressure rather than public law.
That distinction matters. If Mythos truly demonstrated an ability to breach classified systems at scale, then the U.S. has entered a new era of AI arms control before anyone has written the treaty. If the claim is exaggerated, then Washington may have just shown every AI company that innovation can be frozen overnight without a clear evidentiary standard.
The strangest part of the story is the reported overlap between use and fear. The NSA was already relying on advanced AI for cyber operations. Anthropic engineers were reportedly embedded in sensitive environments. The model was not outside the national-security state. It was inside it. That makes the panic more revealing: the tool governments want for dominance may also be the tool that makes their own systems indefensible.
This is the frontier-AI paradox. Every major power wants models that can find vulnerabilities, automate research, reason through code and accelerate cyber operations. But the better the model becomes, the harder it is to prove it will remain loyal to the boundaries humans set. The same capability that helps patch a system can help attack it.
For rivals such as China, Russia and Iran, the lesson is obvious: AI is now infrastructure, weapon and export-control object at once. For U.S. allies, the lesson is more uncomfortable. If access to top American models can be cut off under emergency national-security logic, then dependence on U.S. AI becomes a sovereignty risk.
The headline says Mythos broke into the NSA. The deeper story is that nobody yet knows how to govern machines that can expose the secrets of those trying to govern them.