Anthropic’s Mythos ban and the history of failed export controls

Anthropic was ordered by the White House to restrict export of its AI models Fable and Mythos to anyone outside the United States and to foreign nationals inside the country; Anthropic then pulled both models, which have been unavailable to anyone for a week. The Commerce Department issued an export control directive and the company limited access after being notified, actions the company carried out almost immediately.

What exactly did the White House order and how did Anthropic respond?

The order forced Anthropic to block exports of Fable and Mythos outside the United States and to foreign nationals inside the country; within roughly 90 minutes of being notified Anthropic scrambled to limit access and subsequently pulled both models, leaving them offline for a week. Before the restriction, Mythos had been available only to around 150 vetted companies and government organizations, and Anthropic had launched Mythos in April.

Anthropic’s prior access policy was narrow: the company had positioned Mythos as a limited-access tool for defenders, granting it to about 150 vetted partners to help secure software and services. The immediate pull followed two reported triggers: Anthropic granted access to a South Korean telecom that U.S. officials suspected of links to China, and Amazon CEO Andy Jassy alerted officials after Amazon researchers said they had found a way around safeguards in Fable 5. Anthropic disputes the characterization of that event as a broad defeat of safety work, calling the issue narrow and already patched and disputing the "jailbreak" label.

Why did the administration act now?

U.S. officials flagged concerns after the company’s partner program granted access to a foreign telecom the administration suspected had ties to China, and after Amazon’s leadership raised alarm about an apparent safeguards bypass in Fable 5, prompting the Commerce Department’s export control directive. The combination of a potentially sensitive foreign partner and a reported safeguards bypass convinced officials the models presented an export risk they wanted to contain.

The directive forced an abrupt change in Anthropic’s distribution model, illustrating how quickly an export-control order can curtail access: the company moved from a limited, vetted rollout to a full shutdown of the affected products within a short window after notice.

How does this echo past U.S. export-control efforts against software?

The move mirrors earlier, uneven attempts to curb dual-use software through export controls, from the U.S. probe of Phil Zimmermann in the early to mid-1990s over PGP to later efforts in the early 2010s to expand the Wassenaar Arrangement to cover surveillance and hacking tools. Those prior efforts frequently failed to stop proliferation because countries do not all apply controls the same way and vendors relocate or relicense tools to permissive jurisdictions.

The PGP case became a flashpoint when Phil Zimmermann published PGP’s source code as a printed book during a U.S. criminal investigation, helping end the so-called Crypto Wars and clearing the way for end-to-end encryption used by billions today. In the early 2010s, governments sought to classify spyware as dual-use and require export licenses under Wassenaar, but the treaty depends on member countries’ enforcement and several major spyware vendors relocated or operated from nonadherent states. Some industry moves did produce enforcement results; Germany-based FinFisher shut down in 2022 after prosecutors investigated alleged exports without licenses.

Why it matters

This episode could reshape how U.S. AI companies sell advanced models abroad. If the administration lifts the restriction, it would signal that export controls are a blunt tool that cannot reliably contain frontier capabilities, and that Washington prioritizes U.S. competitiveness. If the administration instead enforces control or requires approvals, American AI companies could face new compliance burdens and limits on foreign customers that could hurt revenue and distribution models.

The Anthropic case tests whether export controls that have had mixed success against encryption and spyware will work against models that are both commercially valuable and widely replicable.

What to watch

Watch whether the administration lifts the restriction or replaces it with a formal approval regime that requires U.S. companies to get government sign-off before serving foreign customers, and whether other labs alter partner programs after this incident. The next concrete signal will be any change in Anthropic’s availability or a formal policy update from the Commerce Department.