China's AI Heist and the FBI's Ghost Town

Anthropic has spent months telling the world its most powerful AI model is too dangerous to release publicly. It turns out that decision may have come a little too late.

According to reporting from Semafor, the White House's move to slap export restrictions on Anthropic's Mythos model was partly motivated by intelligence suggesting a China-linked group had already gotten their hands on it. That's a significant detail, and one that reframes what initially looked like a routine national security precaution into something considerably more urgent.

The model in question — referred to internally alongside a related system called Fable — sits at the top of Anthropic's capability stack. The company has been deliberately secretive about it, positioning Mythos as something too capable and too risky for general availability. Which makes the idea of a foreign state actor getting access not just embarrassing, but genuinely alarming from a geopolitical standpoint.

Here's why it matters beyond the headlines: even if a foreign actor can't fully deploy a frontier model, they can use it as a teaching tool. A technique called model distillation allows a less sophisticated AI — the so-called student model — to learn by observing the outputs of a more powerful one. In other words, access to Mythos, even briefly, could give a rival the building blocks to close the capability gap faster than anyone would like.

The White House hasn't officially confirmed the China angle. Trump advisor David Sacks posted about the export restrictions on X, but framed the concern around jailbreaking vulnerabilities rather than foreign access. Anthropic, for its part, has pushed back on the jailbreaking claims, and a company spokesperson told Semafor that China never came up in government discussions about the export controls.

That's a notable contradiction. Either the White House is keeping its intelligence close to its chest, or the China concern is more speculative than the Semafor report implies. Either way, Anthropic isn't commenting further.

What makes this story sting a little more is that it's not the first time Mythos has slipped through the cracks. Earlier reports revealed that a Discord group had access to the model for roughly two weeks before Anthropic noticed and shut it down. For a company whose entire brand is built on being the safety-first AI lab, that's a difficult pattern to explain away.

The broader takeaway here isn't just about Anthropic. It's about the uncomfortable reality that the most powerful AI systems in existence are being developed and stored by private companies whose security infrastructure is being tested in real time, often by nation-state actors with serious resources and motivation. The race to build the most capable model and the race to secure it are not moving at the same speed.

Somewhere in Huntsville, Alabama, there is a fully functional hospital, a gas station, a hotel, and a row of furnished houses. Nobody lives there. Nobody checks in. The whole place exists so the FBI can tear it apart digitally.

The facility is called the Kinetic Cyber Range, and the bureau quietly opened it last year. This week, for the first time, the FBI released footage giving the public a look inside. At 22,000 square feet, it's less a training room and more a miniature city built specifically to be attacked.

Every building is wired the way it would be in the real world. The hospital has networked medical systems. The gas station has the kind of industrial control hardware that runs pumps and payment terminals. There are home networks inside the furnished houses. And sitting at the center of it all is a small data center running more than 200 servers, ready to be hit with malware, ransomware, or whatever else trainees can throw at it.

The key design feature, and arguably the most important one, is that none of it connects to the outside internet. Whatever gets unleashed inside the Kinetic Cyber Range stays there. That's not a small engineering detail — it's what makes the whole concept viable. You can't run realistic attack simulations on live infrastructure without risking real consequences, so the FBI built infrastructure that looks live but isn't.

The training scenarios are deliberately broad. Agents and analysts can practice forensic investigations on car infotainment systems, trace how an intrusion might spread from a corporate network into employee home setups, or study how a coordinated attack could destabilize a power grid. The point is to expose trainees to the full chain reaction that a real cyberattack can trigger, not just the initial breach.

It's a smart evolution of an old concept. The FBI has run Hogan's Alley — a fake town in Quantico used for physical law enforcement training — since the 1980s. The Kinetic Cyber Range is essentially the same idea updated for an era where the most dangerous criminals don't need to show up in person.

The timing of the public reveal is worth noting. Ransomware attacks on hospitals, water treatment facilities, and energy infrastructure have become a near-weekly headline. The FBI building a dedicated facility to simulate exactly those scenarios signals that the bureau is taking the threat seriously enough to invest in purpose-built infrastructure, not just tabletop exercises.

Whether the training translates into faster, better responses in the field is harder to measure. But the existence of the Kinetic Cyber Range is at least a sign that someone in the federal government understands that defending digital infrastructure requires practicing against it, repeatedly, before the real attack comes.