Last Friday, the White House ordered AI giant Anthropic to immediately halt the export of its powerful Fable and Mythos models to foreign nationals both inside and outside the United States, citing unspecified national security fears. Following the directive, the company swiftly deactivated both systems, leaving them offline for a week. This unprecedented move serves as the first major test of whether Washington can successfully restrict frontier artificial intelligence using export controls—a strategy that has historically yielded highly volatile results when applied to encryption and espionage software.
The High Stakes of the Anthropic Standoff
The resolution of this high-profile conflict will do more than just dictate Anthropic’s access to international markets; it will establish the regulatory blueprint that all future AI laboratories must navigate. Since launching Mythos in April, Anthropic has positioned the model as a highly potent cyber-capability tool capable of causing widespread digital disruption if leaked. Consequently, access was strictly limited to approximately 150 vetted corporations and government bodies, with the intention of allowing defensive teams to secure their infrastructure before malicious actors could acquire equivalent capabilities.
What Triggered the White House Intervention?
The sudden regulatory crackdown was reportedly sparked by two key events. Firstly, Anthropic granted a South Korean telecommunications firm access to Mythos via its partner programme. US intelligence officials raised concerns after identifying the telecom operator—widely identified as SK Telecom—as a company suspected of maintaining ties to China. However, the South Korean firm has vigorously denied these allegations.
Secondly, Amazon CEO Andy Jassy reportedly notified the Biden administration after Amazon’s internal researchers bypassed the safety guardrails of Fable 5. While Anthropic downplayed the incident as a minor, already-patched vulnerability rather than a systemic failure, the political outcome remained unchanged. The US Department of Commerce issued an immediate export-control directive, forcing Anthropic to restrict access to its platforms within a mere 90 minutes of receiving the notification.
The ‘Crypto Wars’: How PGP Defied the State
This is far from the first time a government has attempted to restrict software distribution under the guise of national security. Historically, such measures have repeatedly failed. The most prominent example occurred during the early 1990s, when computer scientists were pioneering encryption methods to protect data traversing the nascent internet. Among these was Pretty Good Privacy (PGP), a software tool that secured data so effectively that intercepted communications were virtually impossible to decode.
var playerInstance_jwplayer_6a7a09eb49e04 = jwplayer( “jwplayer_6a7a09eb49e04” );
playerInstance_jwplayer_6a7a09eb49e04.setup({
playlist: “https://cdn.jwplayer.com/v2/media/lv0GaEwB”,
});
Viewing PGP as a threat to surveillance capabilities, the US Customs Service launched a criminal investigation against its creator, Phil Zimmermann, for violating arms export controls. In a brilliant move of civil disobedience, Zimmermann bypassed the digital export ban by publishing PGP’s entire source code as a physical book. This triggered the famous “Crypto Wars,” which ended when the government dropped the investigation, ultimately enabling the end-to-end encryption protocols used today by billions on platforms like Signal and WhatsApp.
The Wassenaar Arrangement and Spyware Proliferation
Decades later, in the early 2010s, Western-made spyware was discovered targeting political dissidents in the Middle East. This prompted several nations to expand the Wassenaar Arrangement, an international treaty designed to regulate the export of dual-use technologies—software and hardware with both civilian and military applications. By classifying surveillance tools as dual-use, governments aimed to force spyware developers to obtain export licences.
Structural Flaws in Global Treaties
However, the Wassenaar framework has consistently suffered from two fatal flaws. First, major cyber-hubs like Israel—home to some of the world’s most sophisticated spyware developers—are not signatories to the agreement. Second, enforcement relies entirely on the domestic discretion of individual member states.
For years, Italy permitted the notorious spyware firm Hacking Team to export surveillance tools globally, despite documented evidence that these tools were sold to authoritarian regimes to target journalists and human rights defenders. Similarly, other European nations have exhibited lax oversight. Despite repeated scandals, Europe has consistently failed to stop spyware exports to oppressive states, with critics arguing that recent EU regulatory updates do not go far enough. Consequently, companies like Intellexa simply relocated to nations with weaker regulatory frameworks, while others have eyed moves to jurisdictions like Saudi Arabia.
Rare Enforcement Victories
Despite systemic failures, there have been occasional regulatory successes. In 2022, the German spyware developer FinFisher ceased operations following a lengthy criminal investigation. Prosecutors targeted the firm for allegedly exporting surveillance malware to Turkey without a valid licence, where it was deployed against political dissidents.
Can AI Export Controls Actually Work?
As the standoff between Anthropic and the US government continues, Washington faces a difficult choice. The administration may eventually ease restrictions to preserve the global competitiveness of American tech firms—a concession acknowledging that foreign competitors, including China, will inevitably develop equivalent AI capabilities. Alternatively, imposing strict pre-approval mandates for all foreign transactions would saddle US AI developers with heavy compliance burdens, severely impacting their revenue.
If history is any guide, attempting to restrict the flow of digital code through traditional export controls is fundamentally ill-suited to stopping determined global actors from exploiting advanced dual-use technologies.