Anthropic’s Mythos Ban: Why History Suggests AI Export Controls Will Fail

Last Friday, the White House ordered Anthropic to restrict access to its advanced AI models, Fable and Mythos, citing unspecified national security concerns. Within 90 minutes of notification, the company pulled both models from all users—domestic and foreign—marking the first major test of U.S. government efforts to control frontier AI through export restrictions. The move has left industry observers questioning whether this approach can succeed where similar efforts to contain encryption and spyware have largely failed over the past three decades.

The Mythos Ban: What Happened and Why

Anthropic launched Mythos in April, marketing it as a powerful cyber-defense tool capable of identifying software vulnerabilities before malicious actors could exploit them. Access was tightly controlled from the start, limited to roughly 150 vetted companies and government organizations. The goal was to help defenders secure their systems without unleashing the model’s capabilities broadly.

The trigger for the ban appears to have been two events. First, Anthropic granted access to a South Korean telecom company—widely reported to be SK Telecom—which U.S. officials later identified as having suspected ties to China. SK Telecom has denied any connection to China. Second, Amazon CEO Andy Jassy reportedly alerted the administration after Amazon’s own researchers claimed to have found a way around Fable 5’s safeguards. Anthropic disputes the characterization of this as a full jailbreak, calling it a narrow, already-patched issue.

The Commerce Department issued an export control directive, and Anthropic scrambled to comply. The models remain unavailable as of this writing, and the standoff between the company and the administration continues.

A History of Failed Software Export Controls

Governments have attempted to restrict the spread of dual-use cyber technologies for decades, with a track record that is, at best, mixed. The most famous example is the U.S. government’s pursuit of encryption software in the early 1990s. The target was Pretty Good Privacy, or PGP, a program that made email interception virtually impossible. The U.S. Customs Service opened a criminal investigation against PGP’s creator, Phil Zimmermann, for violating arms export controls. Zimmermann fought back by publishing the source code as a printed book, sparking what became known as the Crypto Wars. The investigation was eventually dropped, and end-to-end encryption—now used by billions of people on Signal and WhatsApp—became ubiquitous.

In the early 2010s, researchers discovered Western-made spyware being used against dissidents in the Middle East. In response, several governments expanded the Wassenaar Arrangement, an international treaty limiting the export of dual-use technologies. The idea was to classify surveillance software as dual-use, requiring export licenses for sales abroad. But the arrangement had two inherent weaknesses: major spyware-producing countries like Israel did not sign on, and enforcement was left to individual member states. Italy, for example, allowed Hacking Team to export spyware to oppressive governments despite the company’s track record of targeting journalists and activists. Other European nations have been similarly lax, and sanctioned spyware consortiums like Intellexa have simply moved operations to countries with weaker controls.

There have been isolated successes. German-based FinFisher shut down in 2022 after a multi-year investigation into allegations it sold spyware to Turkey without a license. But these victories are the exception, not the rule.

Why This History Matters for AI

The parallels between encryption, spyware, and today’s AI export controls are striking. In each case, the government sought to restrict a technology it deemed dangerous, only to find that determined actors—whether individual developers, foreign companies, or entire nations—found ways around the barriers. Encryption source code was published as a book; spyware makers relocated to countries with lax rules; and AI labs in China and elsewhere are likely to reach similar capabilities regardless of U.S. restrictions.

The current impasse between Anthropic and the Trump administration could be resolved in one of two ways. The administration could lift the restriction to keep American AI companies competitive globally—a tacit acknowledgment that containment is impractical. Or, it could double down, requiring government approval for all foreign AI sales, a compliance burden that would inevitably slow innovation and dent profits.

Conclusion

The White House’s order to restrict Anthropic’s Fable and Mythos models is a landmark moment in the regulation of frontier AI. But if history is any guide, export controls are unlikely to be the right tool for the job. The encryption and spyware sagas demonstrate that software—unlike physical goods—is inherently difficult to contain. As the standoff continues, the broader lesson remains: trying to lock down dual-use cyber technologies through government mandates alone is a strategy with a poor track record, and AI is unlikely to be the exception.

FAQs

Q1: What exactly did the U.S. government order Anthropic to do?
The White House ordered Anthropic to restrict access to its AI models Fable and Mythos, citing unspecified national security concerns. The company complied within 90 minutes, pulling the models from all users worldwide.

Q2: Why is the history of encryption and spyware relevant to AI export controls?
Previous attempts to control encryption (like PGP) and spyware (through the Wassenaar Arrangement) largely failed because software is easy to distribute across borders, and determined actors find ways around restrictions. The same dynamics apply to AI models.

Q3: Could the U.S. government’s AI export controls work this time?
It is uncertain. The administration could lift the ban to keep U.S. companies competitive, or it could impose permanent licensing requirements. However, historical precedent suggests that containment is difficult, and AI labs in other countries are likely to reach similar capabilities regardless of U.S. restrictions.