Last Friday, the White House directed Anthropic to halt exports of its frontier AI models, Fable and Mythos, citing unspecified national security threats. The order bars the models from being shared with anyone outside the United States, including foreign nationals within the country, and forced Anthropic to pull both products offline within roughly 90 minutes of being notified. Both Fable and Mythos have now been unavailable to any user for a full week, marking the first major test of whether Washington can apply Cold War-style export controls to frontier artificial intelligence the way it has attempted, with mixed results, to contain encryption software and commercial spyware in past decades.
The trigger came from two separate events. Anthropic had granted access to Mythos through its limited partner program to a South Korean telecom widely reported to be SK Telecom, prompting alarm among U.S. officials who believed the company had ties to China. SK Telecom has denied any such connection. Separately, Amazon CEO Andy Jassy reportedly warned the administration that Amazon's own researchers had discovered a method of circumventing Fable 5's safety guardrails. Anthropic has pushed back on the "jailbreak" characterization, describing the issue instead as a narrow vulnerability that was already patched, rather than a wholesale defeat of the model's safeguards. Regardless, the Commerce Department's directive stood.
The restrictions carry high stakes beyond Anthropic itself. Since launching Mythos in April, Anthropic has positioned the model as a powerful cyber weapon with the potential to cause widespread disruption if released too broadly, and had limited access to around 150 vetted companies and government organizations as a defensive measure. How the standoff resolves could establish the regulatory template other AI labs will have to navigate, while determining whether Anthropic retains access to lucrative foreign markets.
The episode echoes a long history of export control efforts that ultimately failed to halt the spread of sensitive technology. The U.S. government's restrictions on PGP encryption in the 1990s, for instance, became a footnote once the code spread freely online. Whether frontier AI, which can be replicated through stolen weights or distilled into smaller open-source variants, can be contained through similar controls remains an open and increasingly urgent question.