The White House last Friday directed Anthropic to halt exports of its frontier AI models Fable and Mythos to any entity outside the United States, as well as to foreign nationals working domestically, citing unspecified national security concerns. Within roughly 90 minutes of receiving the Commerce Department's directive, according to multiple accounts, the AI company pulled both models offline entirely — and they have now been inaccessible to all users, including the approximately 150 vetted companies and government organizations that previously had access, for a full week. The standoff represents the first major test of whether Washington's export-control playbook, historically used with mixed results to contain encryption technology and commercial spyware, can work for cutting-edge AI.
Two incidents reportedly pushed the administration to act. First, Anthropic had granted access to Mythos through its limited partner program to a South Korean telecommunications firm that U.S. officials believed had ties to China — a company widely identified as SK Telecom, which has publicly denied any such connection. Separately, Amazon CEO Andy Jassy alerted the White House after his company's researchers reportedly discovered a method to bypass Fable 5's safety guardrails. Anthropic has pushed back on characterizing that discovery as a "jailbreak," insisting it was a narrow, already-patched vulnerability rather than a wholesale defeat of the model's protections. The Commerce Department's order landed regardless, forcing Anthropic to scramble.
The episode lands against a longer history of governments trying — and often failing — to keep dangerous cyber capabilities bottled up through export restrictions. From the 1990s crypto wars to more recent efforts to rein in commercial spyware vendors like NSO Group, regulators have repeatedly discovered that once a piece of software or cryptographic knowledge exists, controlling its spread is far easier on paper than in practice. Mythos, which Anthropic launched in April and has publicly described in near-apocalyptic terms as capable of wreaking havoc on the internet if released too widely, was already one of the most tightly controlled AI systems in existence, with access limited to a handpicked set of partners meant to harden their defenses before adversaries caught up.
How the Anthropic situation ultimately resolves could ripple well beyond the company's own foreign-market ambitions. If the Commerce Department's directive holds, it will establish a template that other frontier AI labs — OpenAI, Google DeepMind, Meta, and the rest — will need to design their products around, potentially