Ninety minutes. That is all the time the U.S. government gave Anthropic to pull the plug on its most powerful AI models, Fable and Mythos, after a directive from the Commerce Department hit the company’s inbox. For a week, these frontier models—marketed as "doomsday" cyber tools capable of both securing and shattering the internet—have been effectively dark.

This is not just a corporate compliance headache. It is the first high-stakes stress test of whether Washington can treat artificial intelligence like a physical weapon, using the blunt instrument of export controls to keep frontier technology within domestic borders. History, however, suggests the government is fighting a losing battle.

The Precedent of Failure

The U.S. government’s instinct to classify code as munitions is a reflex honed over decades, and it has a track record of spectacular failure. In the 1990s, the target was Pretty Good Privacy (PGP), an encryption tool that intelligence agencies feared would render their surveillance capabilities obsolete. The government treated PGP like a missile system, launching a criminal investigation into its creator, Phil Zimmermann, for violating arms export controls.

Zimmermann’s defense was simple and brilliant: he printed the source code as a book. By asserting that code was speech, he effectively broke the government’s attempt to treat software as a restricted physical good. The "Crypto Wars" ended with the government’s retreat, paving the way for the end-to-end encryption that now secures the private messages of billions of people. The attempt to stop the technology only accelerated its adoption.

The Spyware Trap

If the 90s were about encryption, the 2010s were about surveillance. When Western-made spyware began appearing on the devices of dissidents and journalists, governments turned to the Wassenaar Arrangement—an international treaty designed to regulate dual-use technologies. The logic was that if you force companies to get export licenses, you can stop the flow of hacking tools to authoritarian regimes.

It hasn't worked. The arrangement suffers from two fatal flaws: it is voluntary, and it relies on the discretion of individual nations. Israel, a global hub for the spyware industry, never signed on. Meanwhile, European nations like Italy have historically granted export licenses to companies like Hacking Team, even as those tools were used to target human rights activists. The result is a fragmented, porous regulatory landscape that rewards the most lax jurisdictions.

Why AI Is Different (And Harder)

The current standoff with Anthropic, reportedly triggered by concerns over a South Korean telecom’s ties to China and a potential "jailbreak" discovered by Amazon researchers, highlights the unique volatility of AI. Unlike a static piece of encryption software or a finished spyware package, frontier AI models are constantly evolving. They are not just tools; they are platforms that can be updated, fine-tuned, and accessed remotely.

Anthropic disputes the "jailbreak" narrative, characterizing the Amazon incident as a narrow, patched vulnerability rather than a systemic failure. Yet, the government’s reaction—a total export ban—suggests a lack of faith in the company’s ability to keep the genie in the bottle. If the U.S. cannot trust a single, centralized lab to manage access, it is unlikely that export controls will provide the security it seeks.

What This Means for Developers

For the AI industry, the implications are immediate. If the government establishes a precedent where a single, unverified report of a vulnerability can lead to an overnight export ban, the incentive to build and deploy frontier models in the U.S. diminishes. Developers are already looking at the "Mythos" precedent and asking whether their own infrastructure could be shuttered by a similar administrative order.

If history is any guide, these controls will likely push development into jurisdictions that are less concerned with U.S. national security directives. Just as encryption moved to the open-source community and spyware moved to the most permissive markets, frontier AI will follow the path of least resistance.

Key Takeaways

  • Export controls are reactive: The U.S. government is attempting to apply 20th-century regulatory frameworks to 21st-century software, a strategy that has historically failed to stop the proliferation of code.
  • The "Mythos" precedent: The sudden shutdown of Anthropic’s models signals a new, aggressive era of government intervention in AI development, prioritizing containment over market access.
  • The inevitability of leakage: Because AI models are increasingly accessible via cloud APIs, traditional export controls struggle to account for the borderless nature of digital infrastructure.

The standoff between Anthropic and the Commerce Department is far from over. While the models remain offline, the industry is waiting to see if this is a temporary pause or a permanent shift in the regulatory environment. The next decision point will come when Anthropic attempts to restore access—and whether the government demands a level of oversight that is technically impossible to guarantee.