On a Friday that Anthropic had not anticipated, a government directive arrived with no detailed explanation — only a verbal claim that someone had found a way to make Fable 5, one of the company's most capable AI models, identify software vulnerabilities in ways it was never supposed to. The US government, citing national security, ordered Anthropic to suspend access to both Fable 5 and Mythos 5 for all foreign nationals worldwide, and the company said it would comply by abruptly disabling the models for those users. Hundreds of millions of people use Anthropic's products, making the scale of the disruption exceptionally large. The directive represents something genuinely new in Washington's long-running effort to limit rivals' access to American AI technology.

For years, export controls have focused on the hardware layer — the specialised chips made by companies such as Nvidia that are required to train and run powerful models. Restricting the models themselves, as finished products delivered over the internet, is a different kind of intervention entirely, and one that analysts had long debated in theory but rarely seen practised. Anthropic pushed back firmly, even as it announced compliance. The company argued that the government had provided only "verbal evidence of a potential narrow, non-universal jailbreak" — meaning the vulnerability, if real, would not affect every user or every use case.

It also noted the irony of the timing: as recently as the Wednesday before the order, Anthropic had publicly called for stronger federal oversight of AI models that pose unacceptable risks.