The Mythos Blackout: Why AI Export Controls Are Facing a PGP-Style Crisis
Table of Contents
A Sudden Digital Curtain
Last Friday, the White House abruptly intervened in the rollout of Anthropic’s most advanced AI capabilities. Citing unspecified national security concerns, the administration ordered the AI lab to restrict the export of its Fable and Mythos models to anyone outside the United States, including foreign nationals residing within the country. In a move that underscored the volatility of current regulatory environments, Anthropic pulled the plug on both models almost immediately, leaving them unavailable to all users for over a week.
The blackout represents the first high-stakes test of whether the U.S. government can effectively use export controls to contain “frontier AI.” If the administration succeeds, it could establish a rigid precedent for how all future large-scale models are distributed globally. If it fails, it will be the latest entry in a long history of government attempts to fence in mathematics and code—efforts that have almost always been bypassed by the very nature of the internet.
The Trigger: Geopolitics and Jailbreaks
The crackdown didn’t happen in a vacuum. Since April, Anthropic has positioned Mythos as a double-edged sword: a high-capability system capable of identifying critical software vulnerabilities, but one that could potentially be weaponized for cyberattacks if leaked. Because of this, access was limited to roughly 150 vetted organizations. However, that vetting process appears to have failed.
Two specific incidents reportedly spooked Washington. First, Anthropic granted access to Mythos via its partner program to a South Korean telecom firm—widely believed to be SK Telecom—which U.S. officials suspect has ties to the Chinese government. Though SK Telecom has denied these connections, the suspicion alone was enough to trigger a Commerce Department directive.
Simultaneously, pressure mounted from within the industry. Amazon CEO Andy Jassy reportedly alerted the White House after Amazon researchers discovered a way to bypass the safety guardrails of Fable 5. While Anthropic has dismissed this as a “narrow, already-patched issue” rather than a total system failure, the perception of a “jailbroken” frontier model created a sense of urgency in the Oval Office. The result was a scramble to limit access within 90 minutes of notification.
The Ghost of the Crypto Wars
To understand why this move is viewed with skepticism by digital rights advocates and historians, one only needs to look back at the 1990s. The current standoff mirrors the “Crypto Wars,” where the U.S. government attempted to classify encryption software as munitions.
The most famous casualty of that era was Pretty Good Privacy (PGP). Created by Phil Zimmermann, PGP allowed users to encrypt data, effectively blinding U.S. intelligence agencies. The government responded with a criminal investigation into Zimmermann for violating arms export controls. In a move of strategic defiance, Zimmermann published the PGP source code in a printed book, utilizing the First Amendment to bypass export laws. The government eventually backed down, but the precedent was set: once code is public, it cannot be “un-exported.”
The Failure of the Wassenaar Model
The attempt to control AI also echoes the failures of the Wassenaar Arrangement, an international treaty designed to limit the export of “dual-use” technologies—tools that can be used for both civilian and military purposes. In the 2010s, this was used to target the spyware industry.
The strategy failed for two reasons: inconsistency and mobility. While some nations signed on, others, including Israel, maintained more permissive environments for spyware developers. Meanwhile, companies like the sanctioned Intellexa consortium simply shifted their operations to jurisdictions with laxer oversight, such as Saudi Arabia. Even within Europe, the Italian government previously granted licenses to Hacking Team, despite evidence that the company’s tools were being used to target dissidents and journalists.
The only notable success in this realm was the 2022 shutdown of Germany-based FinFisher, which occurred only after a grueling multi-year criminal investigation. For most, however, export controls functioned as a suggestion rather than a barrier.
The Impasse
As the standoff between Anthropic and the administration continues, the central question remains: can a model be restricted when its weights can be leaked, stolen, or replicated? Unlike a physical chip or a missile component, an AI model is a set of mathematical weights. If the U.S. government manages to keep Mythos within its borders, it may hinder competitors globally, but it also risks isolating American AI labs from the global feedback loops that drive innovation. For now, the world’s most powerful “doomsday machine” remains offline, waiting to see if the government can actually enforce a digital border.
