It was one of the shortest deadlines an AI lab has ever been handed. On Friday, June 12, the US government called Anthropic at around 1:00 p.m. local time (ET) and issued an ultimatum: the company was to shut down access to its newest models, Fable 5 and Mythos 5, within roughly 90 minutes — citing an unspecified national security risk. The formal letter arrived, according to Anthropic, at 5:21 p.m. ET.

Its content: an export control order prohibiting access to both models by “any foreign national” — whether inside or outside the US, and explicitly including Anthropic’s own non-American employees.

An order that technically cannot be split in half: because a user’s nationality cannot be verified in real time on a per-API-session basis, Anthropic’s only option was to deactivate both models entirely for all customers — including the very US citizens who should, in principle, have retained access. Access to all other Claude models remained intact.

Fable 5 was supposed to become Anthropic’s new cash cow

The irony: Fable 5 and Mythos 5 had been unveiled with great fanfare only days earlier. Mythos 5 went to a select group of agencies and companies; Fable 5, the variant equipped with additional safeguards, was made publicly available as “safe for general use.” Both are built on the same foundation as Anthropic’s “Mythos Preview,” which the company itself had classified as too dangerous for public release. It was precisely that earlier warning that now caught up with Anthropic.

Not insignificant from an economic standpoint: Fable 5 was meant to become Anthropic’s new bestseller. In benchmarks the AI model sits well ahead of the competition in some areas — and, incidentally, ahead on price too. At launch, Fable 5 was to be included at no extra charge in the Pro, Max, Team, and seat-based Enterprise plans until June 22. On June 23, Anthropic then planned to remove Fable 5 from those plans — continued use would have required additional usage credits from that point on. Only once sufficient capacity was available did Anthropic intend to bring Fable 5 back as a regular component of the subscription plans, and to do so as quickly as possible.

The trigger: a “jailbreak” of uncertain magnitude

It was apparently Amazon that set things in motion. According to reports from Semafor and Fortune, Amazon researchers used a series of prompts to override Fable 5’s safeguards and got the model to disclose otherwise restricted information about cyberattacks. Amazon CEO Andy Jassy is then said to have raised the matter directly with the administration. Amazon holds a multi-billion-dollar stake in Anthropic and supplies much of its cloud infrastructure — a detail that does not make the episode any less complicated.

The most explosive public appearance came from David Sacks, co-chair of the President’s Council of Advisors on Science and Technology and the White House’s former “AI czar.” In an X post on June 13 he laid out the government’s reading: Amodei had been informed of the jailbreak but had played it down as not serious and refused to fix it. The export control had been imposed “reluctantly”; as soon as the gap was closed, the restriction was to be lifted again — “the ball is in Anthropic’s court.” Sacks at the same time sought to decouple the action from the earlier conflicts between Anthropic and the government: anyone seeing a connection there, he said, was mistaken.

Anthropic firmly disputes this account. In its statement the company calls the bypass “narrow and not universal”: at its core, it amounts to having the model read a codebase and find existing software vulnerabilities — something other publicly available models can do too, with OpenAI’s GPT-5.5 explicitly named. The company says it received only “verbal indications” of a narrowly defined vulnerability, not a single concrete harmful result. And it considers it wrong for a narrowly circumscribed jailbreak to force the recall of a model used by hundreds of millions of people.

The China angle: reports without hard evidence

Over the weekend, a second, more explosive layer settled over the jailbreak dispute. Semafor reporter Reed Albergotti reported that the export decision had been driven, at least in part, by concern that a group linked to China might have gained access to Mythos — with the risk that the model could be reverse-engineered or “distilled.”

The crucial distinction here: this is reporting based on a person familiar with the matter, not a confirmed finding. Anthropic, for its part, pushes back: the White House, it says, did not raise Chinese access at all in the conversations surrounding the Fable jailbreak, and the company blocks access to its products from within China anyway. What remains at this point is an open, unconfirmed question — but one cleanly flagged as unconfirmed by reputable media. It is precisely this gap, as will become relevant shortly, that is now being filled from another quarter with entirely fabricated content.

Who had access to Mythos 5 in China, how, and whether at all, remains open. What is certain, in any case, is that Anthropic’s AI models are rather popular with Chinese companies for so-called distillation. AI distillation (knowledge distillation) refers to a process in which a large, powerful “teacher” model is used to train a smaller, leaner “student” model that mimics its behavior and capabilities as closely as possible — at significantly lower compute and resource requirements.

Anthropic accused the Chinese AI companies DeepSeek, Moonshot (Kimi models), and MiniMax in early 2026 of misusing its Claude models to build their own LLMs. It is quite possible that an attempt was made to continue this practice.

The backstory: the Pentagon dispute that was never fully settled

The current escalation does not come out of nowhere. It is the latest chapter in a months-long power struggle between Anthropic and the Trump administration.

At its heart are two “red lines” that Anthropic drew from the outset: Claude must be used neither for the mass surveillance of US citizens nor in fully autonomous weapons systems that select and engage targets without human intervention. The Pentagon, by contrast, wanted use “for all lawful purposes” — without restriction. You cannot, the defense side argued, allow a private firm to tie the military’s hands in a national emergency.

The negotiations collapsed at the end of February. After Anthropic let a deadline lapse, President Trump ordered federal agencies to stop using Anthropic; the military was granted a six-month transition period. Defense Secretary Pete Hegseth had Anthropic classified as a “supply chain risk” — a label otherwise applied to companies with ties to foreign adversaries. The tone turned personal: Hegseth called Anthropic “self-righteous,” defense CTO Emil Michael attributed a “god complex” to Amodei, and Trump described the company as “radical left” and “woke.”

Anthropic filed suit on March 9 in federal court. With mixed success: federal judge Rita Lin (a Biden appointee) issued a preliminary injunction in Anthropic’s favor on March 26, writing in an unusually sharp 43-page ruling that nothing in the law supported the “Orwellian” notion that a US company could be branded a potential adversary merely for disagreeing with the government. But the appeals court in Washington overturned the stay again on April 8. In May, the Pentagon then closed AI deals with eight major tech companies — Anthropic was not among them. At the same time, the White House recently resumed talks after Anthropic had delivered several technological breakthroughs.

Against this backdrop, an Anthropic delegation led by Amodei met on Monday in Washington with the US Department of Commerce under Howard Lutnick. Further documentation of the alleged problem was among the things expected. Whether Fable 5 and Mythos 5 would be reactivated afterward was initially unclear.

The industry pushes back

Notable is who is taking Anthropic’s side in this dispute. In an open letter at freefable.org, security experts from Nvidia, Zoom, and Mercedes-Benz, as well as former security officials from the US government and Google, called on Secretary Lutnick to lift the restrictions on Fable 5 and Mythos 5. They also demanded that the government commit to an “open, scientific, and transparent” process for future AI risk assessments. To deprive the best defenders of the strongest tools while the other side rapidly arms up, the argument went, is dangerous.

FUD: the myth of the “shadow emperor” of DeepSeek

As soon as “China” and “AI model” appear in the same sentence, disinformation is never far behind. Parallel to the Washington wrangling, a narrative is now circulating that casts CEO and founder Dario Amodei as the secret mastermind behind China’s AI lab DeepSeek.

The text — dressed up as a “two-year investigative report” — claims that Amodei steered DeepSeek from the shadows for years, smuggled architectural ideas from OpenAI to Hangzhou, and founded Anthropic merely as Western cover. The “evidence” offered: a founding idea supposedly sketched on a napkin in a Beijing hotpot restaurant, an anagram game (“D-Seek” as an encrypted “Seeking Dario”), and a “D” read into DeepSeek’s whale logo. The material relies on “obscured financial documents,” old WeChat chats, and anonymous sources — that is, on nothing verifiable.

The real kernel: Dario Amodei did in fact work, as anyone can simply see on his LinkedIn profile, from November 2014 to October 2015 as a Research Scientist at Baidu, at the Sunnyvale location. There he conducted research in Andrew Ng’s team and contributed substantially to “Deep Speech 2,” a speech recognition system that MIT Technology Review counted among the ten most important technological breakthroughs of 2016. That is public, documented, and entirely unspectacular — a top researcher in speech recognition, years before Anthropic existed. Baidu is, moreover, a publicly traded, commercial corporation and has nothing to do with DeepSeek.

That this narrative is gaining traction precisely now is hardly a coincidence. It opportunistically latches onto the real — and legitimate — debate about possible Chinese access to Mythos and tries to turn it into a personal attack on Amodei.

What’s at stake

Beyond anagrams and hotpot anecdotes, what is at issue in Washington is a concrete question of industrial policy: who decides in the future which AI capabilities a US company may ship worldwide — and by what process? The administration has signaled that any future model exceeding Mythos’s performance threshold must pass through the government before release. For an industry whose business model rests on global reach, that is a deep incision.

Critics are already warning that a step like this, of all things, plays into the hands of European and other non-American AI competitors: if US models can go offline at the stroke of a pen, sovereignty over critical infrastructure becomes a selling point. For the European market — and for everyone who depends here on reliable availability — that is the truly relevant lesson of this weekend.