The NSA's Director Says Mythos Broke Into Almost All Its Classified Systems in Hours. The Government's Response Has Been Chaos.

On June 11th, Senator Mark Warner — vice-chair of the Senate Intelligence Committee — said something in public that the public was not really meant to hear so plainly. General Joshua Rudd, who runs both the National Security Agency and the Pentagon’s Cyber Command, had told him that Claude Mythos, Anthropic’s most capable model, “broke into almost all of our classified systems, not in weeks, but in hours.”

Reported by The Economist, that single sentence is the most concrete public datapoint to date on what a frontier AI model can do as a cyber-weapon. The day after Warner repeated it, the US government pulled Mythos — and its constrained sibling, Fable 5 — off the market for every foreign national on Earth. The two events are not as cleanly connected as that timeline suggests. But together they mark the moment American AI policy stopped being a debate about hypotheticals.

What Rudd actually said

The line, as carried by The Economist:

It is worth being precise about what this is and isn’t. Mythos did not “break encryption.” Defeating cryptography — mathematically cracking the ciphers that secure communications — is a different and narrower feat. What is described here is intrusion: finding software weaknesses, chaining them, and moving through a network to reach systems that are supposed to be sealed. That is the capability Mythos has been documented to have since Project Glasswing, where it found a 17-year-old FreeBSD remote-code-execution bug and wrote a working kernel exploit in four hours.

The Economist makes the distinction itself, and uses it to set the stakes. The paragraph before the quote recalls how, from the 1970s to the 1990s, Washington treated strong encryption as a munition — restricting its export, even investigating a developer under arms-trafficking law — before civil-liberties advocates won the right to use and sell it. Then the magazine’s pivot: “Encryption was a potent technology, but narrow in its application. AI is far more powerful and versatile.” The Rudd anecdote is the evidence for that sentence. The old crypto-export panic is the small analogy here.

What the report does not give is method. It does not say how the model was pointed at those systems, who supervised it, or what “almost all” means in practice. A controlled internal test is the only sane reading — an agency chief would not relay an unauthorized breach of his own classified networks to a senator as a capability anecdote — but the specifics are not on the record.

The ban it didn’t quite cause

It is tempting to draw a straight line from the June 11th briefing to the June 12th shutdown: the NSA saw what Mythos could do, and the government slammed the door. The reporting does not support a line that clean. The Economist’s whole thrust — its headline included — is that the blocking of Anthropic was capricious and chaotic.

The official trigger was a jailbreak, not the NSA test. Amazon’s CEO Andy Jassy — whose company holds a large Anthropic stake — reportedly alerted the administration that it had found a way to override Fable 5’s guardrails. White House adviser David Sacks posted that the government asked Anthropic to “fix the jailbreak or de-deploy” Fable, and that “Dario refused.” Anthropic’s own account mentions no refusal, only an instruction to block foreign nationals over an unspecified national-security threat. The company says the jailbreak surfaced merely a “small number of previously known, minor vulnerabilities,” and vulnerability researcher Katie Moussouris, who reviewed the underlying paper, said the techniques were aimed at helping defenders.

And there is the politics. The Economist floats that export controls were “a convenient way to target Anthropic,” noting that President Trump has called the company an “out-of-control Radical Left AI company,” and that on June 13th the secretary of war, Pete Hegseth, boasted he had “kicked Anthropic out” of the Pentagon months earlier. The legal mechanism, revealed days later in the Lutnick letter, was the export-control doctrine of “deemed export” — not a citation of the NSA result at all.

So the honest shape of the story is two things at once: a genuine, frightening capability, and a government response to it that has been erratic and possibly political. The Rudd quote is the clearest signal of why anyone should be alarmed. It is not the stated reason the door was shut.

Why this dwarfs the encryption wars

The Economist reaches past the crypto analogy to a heavier one: nuclear technology. In 1946 the McMahon Act severed American nuclear cooperation even with close allies. The fear now is similar — that a decisive lead could be lost. The magazine notes that the US advantage could become “unassailable” if a lab cracks recursive self-improvement, with models writing better versions of themselves, while China sits perhaps a year behind on chips.

That framing is why the cutoff is total. The June 12th order exempts no one — not even the Five Eyes intelligence allies (Australia, Britain, Canada, New Zealand). Britain’s AI Security Institute, the leading body in the world for testing and jailbreaking new models, was locked out alongside everyone else. “After a lesson this clear,” former British security minister Tom Tugendhat told the magazine, “every nation will be asking what they need to achieve sovereignty.”

Whether the policy even works is a separate question. A former FBI cyber official notes that hackers buy American identities on the black market to reach restricted models; Anthropic already bars Claude in China, yet Chinese users get in anyway. And the oversight is thinning at the worst moment: a June 2nd executive order asked labs to give the government pre-release access to their top models, even as the Centre for AI Standards and Innovation — the body that vets frontier models for dangerous capabilities — was instructed to stop publishing public reports.

One thread to watch

There is a striking, less-confirmed coda. Separate reporting holds that the NSA is now operating Mythos in its own cyber missions, with Anthropic engineers embedded at the agency — the same model deemed too dangerous for a foreign national to touch, run by US cyber command. It would be consistent with Rudd leading both the NSA and Cyber Command. But it does not appear in the Economist briefing and traces only to thinner reporting, so it belongs in the “watch this” column, not the established record.