The Mythos Ban: Why AI Export Controls Are Repeating the Failures of the Crypto Wars
Table of Contents
The sudden blackout of Mythos
Last Friday, the digital availability of Anthropic’s most potent models vanished. Citing unspecified national security concerns, the White House ordered the AI lab to restrict the export of Fable and Mythos to any entity outside the United States, including foreign nationals operating within U.S. borders. The response from Anthropic was swift: the plug was pulled on both models within approximately 90 minutes of notification, leaving a void for the small group of vetted users who previously had access.
This move represents the first high-stakes test of the U.S. government’s ability to treat frontier AI as a controlled munition. For months, Anthropic had positioned Mythos as a double-edged sword—a “doomsday cyber machine” capable of identifying critical software vulnerabilities at an unprecedented scale. To mitigate this, access had been limited to roughly 150 vetted organizations, primarily intended to give defenders a head start in securing infrastructure before adversaries developed similar capabilities.
The catalysts for intervention
Two specific triggers reportedly pushed the administration toward a hard ban. First, Anthropic granted access to Mythos via its partner program to a South Korean telecommunications firm—widely believed to be SK Telecom. U.S. officials grew alarmed over alleged ties between the telecom provider and Chinese interests, despite denials from the company.
Simultaneously, internal pressure mounted from the private sector. Amazon CEO Andy Jassy reportedly alerted the administration after Amazon researchers discovered a method to bypass the safeguards of Fable 5. While Anthropic has pushed back against the “jailbreak” narrative, characterizing the incident as a narrow, patched edge case rather than a systemic failure of its safety architecture, the political optics remained damaging.
A history of leaked boundaries
To understand why this ban might be futile, one only needs to look at the 1990s. The U.S. government once viewed Pretty Good Privacy (PGP) encryption as a threat to national security, fearing that end-to-end encryption would blind intelligence agencies. The Customs Service launched a criminal probe into PGP creator Phil Zimmermann for violating arms export controls. In a move that became a legendary act of digital defiance, Zimmermann published the PGP source code as a printed book, arguing that the distribution of text is protected speech.
Zimmermann eventually won, and the “Crypto Wars” effectively ended with the realization that mathematics cannot be legislated. PGP paved the way for the modern encrypted world, from Signal to WhatsApp. The lesson was clear: once a piece of logic or code is released into the wild, export controls are merely speed bumps.
The spyware loophole
More recently, the international community attempted to curb the spread of offensive cyber tools through the Wassenaar Arrangement. By classifying surveillance and hacking software as “dual-use” technologies, the treaty forced developers to seek export licenses.
However, the arrangement suffered from systemic leakage. Israel, a global hub for spyware development, did not adhere to the agreement. Meanwhile, European nations often applied the rules inconsistently. Italy, for instance, granted licenses to Hacking Team despite evidence that the company was selling tools to oppressive regimes to target journalists. When pressure mounted in Europe, firms like Intellexa simply migrated operations to jurisdictions with laxer oversight, such as Saudi Arabia.
The frontier AI dilemma
Anthropic now finds itself in the center of this historical pattern. Unlike physical hardware—like the H100 GPUs that the U.S. already restricts via Commerce Department rules—AI models are essentially weights and parameters. If a model’s weights are leaked or if a “jailbreak” allows for remote exploitation, the border becomes irrelevant.
As the standoff between the Trump administration and Anthropic continues, the outcome will likely set the precedent for every other frontier lab. If the government succeeds in maintaining a total blackout, it may signal a new era of “digital sovereignty” where AI is treated as a state secret. If the models eventually leak or the administration buckles under economic pressure to maintain U.S. competitiveness, it will be another chapter in a long history of governments trying—and failing—to put a leash on the internet.
