A US export control directive forced Anthropic to restrict access for foreign nationals — and the company says the downstream effect was pulling Fable 5 and Mythos 5 from every customer, including Americans.
On June 13, 2026, the US government ordered Anthropic to restrict access to its newest AI models for foreign nationals, including its own foreign-national employees. In compliance, every customer worldwide — including Americans — lost access.
The order is an export-control directive, a legal tool normally used to stop sensitive hardware and dual-use technology from reaching foreign adversaries, and the rationale is a national-security finding that one of the models can be jailbroken — a way to trick a model into ignoring its safety rules.
Anthropic disputes the technical basis. In a post on X, aggregated by Latent Space's daily AI news roundup, the company said the order rests on verbal evidence of a "narrow, non-universal jailbreak," and that comparable capabilities are already "widely available" in OpenAI's GPT-5.5 and other frontier systems. The company added its standard line that it "steers clear of commenting on politics and policy."
The models in question are Fable 5 and Mythos 5, released by Anthropic roughly three days before the suspension. The Latent Space roundup frames the suspension as a notable departure from prior Anthropic friction with Washington, which had previously affected only US government employees and vendors.
The scope is what makes this a precedent rather than a familiar export-control story. The directive applies to all customers worldwide, which the roundup notes is a much wider reach than the typical "foreign nationals only" language of past US export actions. The order also covers Anthropic's own foreign-national employees. In effect, the government has used an export-control tool whose practical effect was taking a US-built software product off the US market as well as the foreign market.
That is the contest the story now turns on. One side is a US government claim, delivered through an export-control mechanism and not yet published in full, that a jailbreak of a frontier model rises to a national cybersecurity risk. The other side is Anthropic's technical rebuttal: the jailbreak is narrow, the capability is not unique to these models, and pulling them for every customer, including US customers, is the wrong remedy. Both claims are unresolved, and the resolution will set the rule for every lab that ships a model the US government decides is a security asset or a security risk.
The next signals to watch are three. First, the primary text of the directive, which has not yet been published and may narrow or broaden the "national cybersecurity" rationale. Second, an independent press account, since so far the only direct company statement is the single X post surfaced by the Latent Space roundup. Third, the response from the open-source and open-weights community, which has argued that pulling a model over a contested jailbreak claim chills open publication and leaves the field to closed labs.