For three days, Anthropic’s newest AI model looked like the future of artificial intelligence. Then Washington pulled the plug.

Anthropic’s Fable 5 and Mythos 5 models were supposed to mark a new stage in the AI race: systems powerful enough to autonomously hunt software flaws, reason across massive codebases, accelerate drug discovery, and handle long, complex tasks that earlier models could not. Instead, they have now become something even bigger, the first major test of whether the U.S. government will treat frontier AI models not simply as products, but as national-security assets.

The shutdown came after the U.S. government issued an emergency export-control directive barring access to Fable 5 and Mythos 5 by any foreign national, whether overseas or inside the United States. That reportedly included Anthropic’s own foreign-national employees. Because Anthropic could not practically comply while keeping the models available to ordinary customers, the company disabled access for everyone.

The result was immediate and stunning, one of the world’s most valuable AI startups had to yank its most advanced models just days after launch, at the exact moment it was preparing for a historic IPO.

What makes this story so explosive is the chain of events behind it. According to multiple reports, the crackdown was triggered after Amazon raised alarms with senior U.S. officials about Anthropic’s newest model. Amazon researchers had reportedly used Fable 5 to surface software vulnerabilities through a series of prompts, warning the government that the model could provide information useful to cyberattackers.

That warning moved fast. Amazon CEO Andy Jassy reportedly discussed the matter with senior Trump administration officials, including Treasury Secretary Scott Bessent. White House officials then met to decide how to respond, security researchers tested Amazon’s claims, and officials pressed Anthropic to fix the issue or take the model down. When the company did not voluntarily pause the model, the Commerce Department moved through export controls.

By late Friday, users were losing access.

Anthropic’s response was unusually direct. The company said the government’s letter did not provide specific details about the national-security concern. It said it had received only verbal evidence of what it described as a narrow, non-universal jailbreak, essentially involving the model reviewing code and identifying software flaws. Anthropic argued that the vulnerabilities were minor, already known, and discoverable through other public models.

That is the core dispute. The government appears to see Fable and Mythos as a new class of cyber-capable AI system that could give foreign actors dangerous leverage. Anthropic says the cited example does not show anything uniquely dangerous about its model and that the same kind of defensive security work is already being done every day by legitimate cyber teams.

The models at the center of the fight are not ordinary chatbots. Mythos was Anthropic’s more sensitive system, initially held back from broad release because of its ability to find and exploit software vulnerabilities. Under Project Glasswing, Anthropic gave selected cyberdefenders and critical infrastructure partners access to Mythos Preview to help find and patch flaws in important systems before hostile actors could use similar AI capabilities offensively.

That project was not small. Partners included major technology and infrastructure names such as Amazon Web Services, Apple, Cisco, CrowdStrike, Google, Microsoft, NVIDIA, Palo Alto Networks, JPMorganChase, the Linux Foundation, and others. Anthropic said Mythos Preview had already found thousands of high-severity vulnerabilities, including flaws in major operating systems and browsers.

Fable 5 was supposed to be the broader public compromise: a Mythos-class model with stricter safeguards. Anthropic marketed it as a major leap in coding, science, finance, vision, and long-horizon autonomous work. It said Fable could operate across huge codebases, use memory more effectively, interpret complex charts and tables, and perform advanced scientific reasoning. Mythos 5, meanwhile, was the same underlying model with certain safeguards lifted for trusted partners, especially in cybersecurity.

That distinction now matters. Washington did not just restrict Mythos, the more sensitive version. It also moved against Fable, the supposedly safer public release.

This is why the order is so consequential. For years, U.S. AI export controls focused mainly on chips, semiconductor equipment, and model weights. The logic was simple: stop adversaries from acquiring the compute needed to train frontier systems. This case shifts the battlefield from hardware to access. It treats use of a model itself as something that may need to be controlled by nationality.

That is a much more aggressive theory of AI governance. If a foreign national working inside a U.S. company cannot access a model without triggering export-control concerns, frontier AI starts to look less like software-as-a-service and more like a classified industrial capability. That has huge implications for labs that rely on international talent, global enterprise customers, cloud platforms, and allied-nation cooperation.

It also creates a brutal practical problem, how does an AI company verify nationality at model-access scale? Location is easy. Citizenship is not. If the rule is enforced literally, companies may need new identity systems, licensing workflows, internal access controls, cloud-region restrictions, and employee firewalls around frontier models. The ripple effect could be enormous.

There is also a deeper strategic contradiction. The Trump administration had just created a voluntary framework for advanced AI testing, explicitly saying it was not building a mandatory licensing regime for new models. But the Anthropic order functioned, in practice, like a licensing regime. Axios quoted one person familiar with the matter calling it exactly that: a de facto licensing system.

That does not mean every AI lab is next. The Information reported that the government is unlikely to extend the Anthropic export control to other AI companies, with an official saying the directive was specific to Anthropic’s response to the government’s concerns. But that may be cold comfort. The precedent is now visible: if Washington decides a model crosses a national-security line, it can move fast, forcefully, and with limited public explanation.

But even if access is restored, the old world is gone. Frontier AI companies now know that a model launch can become a national-security event within hours. Investors now know that trillion-dollar AI valuations carry political risk. Foreign users now know that access to the most powerful American models may not be guaranteed. And governments now know that export controls can reach directly into the AI product layer, not just the chip supply chain.

Fable was live for only days. The precedent may last for years.