An AI Broke Into Almost All of the NSA's Classified Systems in Hours

The model was Mythos, reported to be Anthropic's most powerful system. The senator was Mark Warner of Virginia, vice chair of the Senate Intelligence Committee. The general was Joshua Rudd, who runs both the NSA and U.S. Cyber Command. In Warner's account, first reported by The Economist and picked up across the trade press, Rudd told him what the model did.

Mythos broke into almost all of our classified systems, not in weeks, but in hours.

Gen. Joshua Rudd, head of the NSA and U.S. Cyber Command — as relayed by Sen. Mark Warner and reported by The Economist, June 2026

That sentence traveled the way you would expect. "AI hacks the NSA" is a headline that writes itself. It is also not quite what happened, and the gap between the viral version and the real one is worth being exact about, because the real one is the one that should keep you up.

What actually happened

No rogue machine turned on its makers and stormed the nation's secrets. The break-in took place inside an authorized red-team exercise: a controlled security test in which the agency pointed the model at replicas of its own classified environments and told it to find the way in. That is a normal, sober thing for a spy agency to do with a powerful new tool. It is the responsible version of events.

The caveats piled up quickly. A government official later said Warner may have compressed the story in the retelling, and that the "hours, not weeks" line referred specifically to the red-teaming context rather than a live network. Neither the NSA nor Cyber Command has independently confirmed his account. Take all of that on board. Grant every correction. Assume the most reassuring reading survives.

What is left after the deflation is still this. A model you can rent, aimed at stand-ins for the most secure systems this country knows how to build, found the flaws and chained them together into almost total access in a single afternoon. A human red team of the world's best operators does not do that in an afternoon. This is the calm, officially clarified, nothing-to-see-here version of the story. Sit with how little calm is actually in it.

The response is the tell

You can learn a lot about how serious something was by watching what serious people do next. They did not shrug.

The day after the test, the Commerce Department moved to bar every foreign national, including Anthropic's own non-citizen employees, from using its two most capable models, Mythos and Fable. Anthropic did not litigate the technicality. It pulled both models from the market worldwide, for everyone, saying it could not practically enforce a citizenship test on who gets to log in. Read that back. A company took its flagship product, the thing it exists to sell, and switched it off for the entire planet inside a day. You do not do that over a capability you believe is safely under control.

And the NSA test was not a one-off parlor trick. Through a classified evaluation program, Mythos has reportedly surfaced thousands of software vulnerabilities, among them a flaw that had been sitting undiscovered in OpenBSD, an operating system chosen by security professionals precisely because it is hard to break, for twenty-seven years. Twenty-seven years of expert human eyes missed it. The model read the code the way you read a paragraph, and it does not get tired, and there is only ever going to be more of it.

This is the slowest it will ever move

This should alarm anyone paying attention. The labs are racing toward superintelligence and the pace is only accelerating. A year ago, nobody outside a frontier lab had watched a model do what Mythos just did. Six months from now there will be systems that make Mythos look slow, and six months after that the comparison will embarrass us again.

The reason it compounds is not mysterious. A model good enough to find the holes in the world's software is a model good enough to help design its own successor, which is the whole mechanism behind recursive self-improvement. Research that speeds up research does not add. It multiplies. The capability curve does not care that we are not ready for it, and it will not idle at the curb while the law catches up. The law is not close.

I'm done softening it

So I am going to stop choosing my words so carefully, because we are out of time for careful. I have spent a long while wording all of this gently, so that reasonable people would keep listening and not file me under doom. That caution has become a liability. It buys me a reputation for being measured while the thing I am measured about gets built anyway.

A likely outcome of this race, if we do not stop it, is that we lose control of what we are building and it kills us. Not "disrupts the labor market." Not "strains our institutions," though it will do those too, on the way. Kills us. Everyone. Everyone you love, and everyone they love, and the whole quiet human future that was supposed to come after.

I know exactly how that reads on the page. Here is why I will write it anyway, and why it is not the private terror of a fringe. In the largest survey of AI researchers ever conducted, with nearly 2,800 people who build these systems for a living, between 38 and 51 percent put the odds of advanced AI causing outcomes as bad as human extinction at one in ten or higher. Not one in a million. One in ten, from the builders themselves. You would not let your child board a plane with those odds. We are strapping in the whole species.

Geoffrey Hinton won the Turing Award for the methods that made all of this possible, then left Google so he could say what he thinks without a corporate minder in the room. He puts the chance that AI drives humanity to extinction at 10 to 20 percent. He is not the most frightened person in the field. Plenty of the people closest to the work put the number higher, and the estimates tend to climb, not fall, the nearer the expert sits to the frontier.

Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.

Statement on AI Risk, Center for AI Safety (2023)

That single sentence was signed in 2023 by hundreds of the most cited researchers alive and by the chief executives of the very labs racing to build the technology, Anthropic among them. Read it once more and hold what it means. The people building this put its power to end humanity in the same breath as nuclear war, and then went back to their desks and built it faster. That is not my framing of the situation. It is theirs, in their own words, with their own names attached.

But you can do something about it

And this is the part that matters most, because quiet despair is just another way of doing nothing, and doing nothing is how we lose by default. You are not a spectator here.

Mythos got pulled off the shelf because a government pushed, and the government pushed because the pressure reached it. That is the entire machine. Governments write the rules, and governments answer to people who refuse to be quiet. So refuse.

• Tell your representatives, in plain words, that you want binding limits on frontier AI and that you will vote on it. If you have never done it, our tool to write your representative makes the first message easy.
• Back an enforceable international treaty on the most dangerous systems, the way the world once decided that nuclear and biological weapons were too dangerous to leave to whoever moved fastest. It is exactly the work our plan is built around.
• Support the people doing this full time. Add your name below, give what you can, and do it this week, not after the next headline that is worse than this one.

The people building this have told you, on the record and under their own names, that it might kill everyone. An AI just walked through almost every classified wall the country has, in a single afternoon, and the official response was to pull it off the market before the weekend. Stop waiting for a warning clearer than the ones already sitting in front of you. There is not going to be one.