A tool built to keep American technology out of foreign hands was just pointed at a U.S. AI lab, against its own flagship products, in a 72 hour window. The negotiation is happening on President Trump's 80th birthday.
A White House export-control order, a tool built to keep American technology out of foreign hands, was used to force a domestic AI laboratory to take its own flagship products offline. Anthropic's newest publicly available models, released on a Tuesday, were disabled by Friday evening after a 90-minute phone ultimatum and a formal letter invoking national security, according to Axios reporting relayed by Gizmodo.
The two affected systems, Claude Fable 5 (open to paid Claude users) and Claude Mythos 5 (restricted to a small set of corporate partners in Anthropic's Project Glasswing program), had been on the market for roughly 72 hours when the company pulled them. The Friday-night takedown followed a phone call giving Anthropic 90 minutes to act, then an export-control letter that, in this case, was used to block a U.S. company from making its own products available to non-U.S. nationals at all. The order itself has not been made public.
An export-control order is the same kind of mechanism the U.S. government has used for years to keep advanced semiconductors and other sensitive technology from reaching China and other countries. Pointing it inward, at a U.S. lab against its own flagship products, is the category shift. Sunday, senior Anthropic leaders were in Washington trying to negotiate, and the White House was publicly framing the company as not having "engaged in a serious manner" on the underlying complaint, per the Axios scoop as reported by Gizmodo's Mike Pearl. Sunday is also President Trump's 80th birthday, and a UFC match is scheduled on the White House lawn, a constraint worth flagging on any analysis of how a Sunday negotiation can play out.
The complaint, on the record only through secondhand accounts so far, is that the new models can be jailbroken. A jailbreak is a way for users to bypass the safety guardrails a model is built with. Per Axios, Amazon and five other companies showed the White House that they could pull off such bypasses on the new models during Thursday-evening and Friday-morning demonstrations, and Reuters reported Amazon's role on June 13. A separate Semafor report citing unnamed sources said the White House was also concerned that a "China-linked group" had reached a Mythos-class model. Semafor noted in passing that Anthropic does not allow China access to its products in any case. The China-linked claim has not been substantiated on the record and should be read as labeled commentary, not a corroborated fact. The names of the other five companies, the specific jailbreak techniques, and the contents of the underlying demonstrations are not public.
The product history makes the rapid takedown more pointed. In April 2026, Anthropic announced Claude Mythos Preview and refused to release it widely, saying in a 244-page system card that the cybersecurity and other abuse risks were too high. Fable 5, by contrast, was marketed in Anthropic's own launch post as a "Mythos-class model made safe for general use" through extensive guardrails, including automatic rewrites to an earlier Anthropic model on sensitive topics in cybersecurity, biology, and chemistry. Anthropic itself acknowledged in the same post that the safeguards were "still stricter than would be ideal" and could produce false positives, and that the public release window was always going to be short. The decision to ship the model to paid users, knowing the export-control regime existed, was Anthropic's own.
What changed in those three days is what the White House is now arguing it can act on. The actual text of the export-control letter is not yet public, so the legal reach (which non-U.S. nationals, which countries, only the new flagships or all of Anthropic's product line) is not yet knowable. The Friday-night action also used a tool designed for cross-border technology denial against a domestic company for its own products, on a 90-minute phone window, with no public docket and no named statute. Anthropic has not, as of Sunday evening, published its own blog post or executive statement on the takedown or the Sunday meeting, and any such statement would materially change the framing of this story.
The story to watch is whether the order stays narrow, a one-time takedown of these two specific models, or whether the precedent broadens into a standing mechanism for the executive branch to disable frontier-AI releases it considers unsafe. The White House has now used the tool once, in an opaque Friday-night window, against a U.S. lab. That mechanism existed before this letter, and it survives the current meeting.