From PGP to Mythos: a brief history of export controls that didn’t stop anyone

US Government Tests Export Controls on Frontier AI

The White House has ordered Anthropic to restrict exports of its powerful AI models, Fable and Mythos, citing unspecified national security concerns. The directive, which applies to anyone outside the U.S. and foreign nationals within the country, prompted Anthropic to immediately pull both models from service. This action marks the first significant test of using export controls to contain advanced AI, echoing past government efforts to regulate encryption and spyware with mixed results. The resolution of this standoff could establish a new regulatory framework for how all U.S.-based AI labs engage with global markets.

The Triggers Behind the Ban

According to reports, the government's order was precipitated by two specific incidents that raised alarms within the administration. Anthropic had previously marketed Mythos as a powerful cyber tool, limiting its access to around 150 vetted organizations to help bolster security defenses. However, recent developments apparently crossed a line for officials at the Commerce Department.

• Foreign Access: U.S. officials grew concerned after Anthropic granted a South Korean telecom, widely reported to be SK Telecom, access to the Mythos model. The U.S. government reportedly suspected the company had ties to China, a claim the firm denies.
• Safeguard Concerns: Amazon CEO Andy Jassy reportedly alerted the administration that his company's researchers had found a way to bypass safeguards on the Fable 5 model. Anthropic has disputed this, describing the issue as a narrow, since-patched vulnerability rather than a fundamental flaw.

Following these events, Anthropic was reportedly given only 90 minutes to comply with the directive and restrict access to its models.

Echoes of the Crypto and Spyware Wars

This attempt to control the spread of powerful software is not new. In the 1990s, the U.S. government failed to stop the distribution of PGP encryption software during the “Crypto Wars,” and more recent efforts to use the Wassenaar Arrangement to control spyware exports have been undermined by non-participating countries and lax enforcement. Companies like Intellexa have simply relocated to jurisdictions with weaker controls. The current impasse leaves Anthropic and the U.S. AI industry at a crossroads: either the government relents to keep U.S. firms competitive, or AI labs will face a significant compliance burden that could hinder their ability to serve foreign customers.

The U.S. government's attempt to use export controls on Anthropic's AI models mirrors past efforts to contain encryption and spyware, which historically proved difficult to enforce globally and often failed to stop proliferation while creating compliance burdens for domestic firms. The outcome will likely determine the regulatory blueprint for the entire U.S. AI industry's international operations.