US export-control directive took Anthropic's Claude Fable 5 and Mythos 5 offline worldwide

A Friday directive and a global shutdown

At 5:21 p.m. Eastern on Friday, June 12, Anthropic received a directive from the US Commerce Department. By nightfall, according to a detailed account published on dev.to, two of the company's most capable models — Claude Fable 5 and Claude Mythos 5 — had gone dark for every user in the world.

The order was an export control: Anthropic had to prevent any foreign national from accessing the two models, inside or outside the United States, including its own foreign-national employees. With no reliable way to verify every user's nationality in real time, the company suspended both models for everyone rather than risk noncompliance.

The controls held until June 30. On July 1, Fable was restored globally. Mythos returned on a narrower path: after government approval on June 26, access went to a set of US organizations while Anthropic kept coordinating with the government over broader access.

The security finding behind the order

Fable and Mythos share one underlying model, per the dev.to essay. Fable shipped with strong safeguards for general use; Mythos, with fewer guardrails, went only to a small group of defensive-cybersecurity partners. Anthropic has described Mythos as better at finding and exploiting software vulnerabilities than any other model and all but the most skilled human experts.

The finding behind the June directive involved Fable, not Mythos: Amazon researchers prompted it to identify several software vulnerabilities and, in one case, to produce code demonstrating how one could be exploited.

Anthropic's counter-account, as the essay reports, was that less capable models could identify the same vulnerabilities and that every other model it tested reproduced the exploit demonstration. It called the task routine defensive-security work that its safeguards had blocked out of caution. The author counters that Anthropic is an interested party whose interpretation deserves testing against an independent technical standard — precisely what was missing.

The directive arrived with no published severity test, no disclosed threshold and no graduated response. The models went from released to globally unavailable in hours, despite thousands of hours of pre-release red-teaming by Anthropic, the US government, the UK's AI Security Institute and outside organizations. Afterward, Anthropic trained a more targeted classifier — tested by the Commerce Department's Center for AI Standards and Innovation — and proposed a common framework for judging jailbreaks, earlier government access and resources for joint research.

The Pentagon blacklist in the background

The June order did not land in a vacuum. For months, Anthropic and the Pentagon had negotiated over military access to Claude. The Pentagon wanted contractual permission for "all lawful uses"; Anthropic supported government use but insisted on excluding mass domestic surveillance of Americans and fully autonomous weapons, arguing current models were not reliable enough to select and engage targets without human control. Those limits were contractual only — once Claude Gov was deployed, Anthropic could not enforce them technically and had no visibility into how the government used it.

On February 27, after Anthropic refused to remove the limits, President Trump ordered every federal agency to stop using its technology. Defense Secretary Pete Hegseth designated Anthropic a supply-chain risk and barred military contractors, suppliers and partners from commercial activity with it. Trump called Anthropic a "radical left, woke company"; Hegseth accused it of arrogance and betrayal. The authority invoked, Section 3252 of Title 10, is built around the risk that an adversary will sabotage or covertly subvert a national-security system — a poor fit, the essay argues, for a public dispute over contract terms.

The courts push back

Anthropic challenged the measures in federal court. On March 26, Judge Rita Lin issued a preliminary injunction; on August 27 she granted summary judgment on Anthropic's First Amendment, due-process and administrative-law claims. The record behind the designation rested mainly on a four-page memorandum written after two of the three challenged actions. The government abandoned its claim that Anthropic kept backdoor access to deployed models and conceded they were no riskier than comparable black-box models; what remained was distrust rooted in Anthropic's public criticism.

Lin found the designation violated the statute and called the broader measures "illegal and baseless," while noting the Pentagon remains free to choose another vendor. A parallel D.C. Circuit case over a procurement determination under Section 4713 of Title 41 was denied interim relief in April without a merits ruling and remained open as of August 31. The essay cautions that none of this proves the June export directive was retaliation; the two actions rested on different legal processes and different claimed risks.

Why it matters

A model that can vanish worldwide within hours, under a company-specific determination based on facts no customer could read in advance, is more than a technical dependency. Once a model is woven into a product or an operating process, access to it becomes part of the architecture — in the essay's framing, a bet on unpublished policy. The government's concern was real: a model that materially lowers the cost of finding and exploiting vulnerabilities can affect national security even when ordinary users have harmless intentions. But emergency action for a reproducible security problem does not require governing indefinitely by surprise. The United States needs to regulate frontier AI, and it needs companies, researchers, investors and customers to know the rules before they are enforced. June, on this account, showed how far apart those two needs remain.