Washington Just Switched Off Anthropic’s Most Powerful Models

A federal directive forced Anthropic to disable Fable 5 and Mythos 5 for every customer on the planet. The official reason is a jailbreak. The fuller story is a feud that has been building for months.

On Friday evening, June 12, the most advanced commercial AI models in public release went dark. Not because of an outage. Not because of a safety scare from inside the lab. Because the United States government told Anthropic to shut them down.

At 5:21pm Eastern, Anthropic received an export control directive citing national security authorities. The order instructed the company to suspend all access to Fable 5 and Mythos 5 for any foreign national, whether inside or outside the United States, and including Anthropic’s own foreign national employees. There is no clean way to enforce a restriction like that on a live product used by hundreds of millions of people. So Anthropic did the only thing compliance allowed. It pulled both models for everyone.

Every other Anthropic model stays online. Opus, Sonnet, and Haiku, the workhorses behind most day to day Claude usage, are untouched. The blackout is surgical, aimed squarely at the company’s frontier tier.

The Stated Reason: A Jailbreak

According to Anthropic, the government’s concern is a method for bypassing Fable 5’s safeguards. The company says it reviewed a demonstration of the technique, and what it found was underwhelming. The exploit, as Anthropic describes it, amounts to handing the model a codebase and asking it to find and fix software flaws. The vulnerabilities it surfaced were minor and already known. More to the point, Anthropic argues that other publicly available models can find the same flaws without any jailbreak at all, naming OpenAI’s GPT-5.5 directly.

This matters because of how Anthropic built Fable in the first place. The company’s whole pitch was defense in depth. It conceded openly at launch that perfect jailbreak resistance is not currently possible for anyone in the industry. Instead of pretending otherwise, it leaned on layered safeguards, aggressive monitoring, and a thirty day data retention policy designed to catch and shut down attacks fast. Before release, Anthropic says Fable was red teamed for thousands of hours by the US government itself, the UK AI Safety Institute, outside firms, and internal teams. None of them found a universal jailbreak, the kind that broadly unlocks dangerous capability. What the government is reportedly pointing to is the narrow kind, and a narrow one that, by Anthropic’s account, produced nothing genuinely dangerous.

In plain terms, Anthropic’s position is this. If a minor, narrow vulnerability that every frontier model shares is grounds for yanking a deployed product, then no frontier model can ever ship again.

The Backdrop Nobody Should Ignore

Here is the part that turns this from a compliance headline into a real story.

This is not Anthropic’s first collision with Washington this year. Back in February, contract negotiations between Anthropic and the Department of Defense, which the current administration has renamed the Department of War, collapsed. The sticking point was not money, although a defense contract worth up to $200 million was on the table. It was use. The Pentagon wanted an “all lawful purposes” standard that would let the military use Claude for effectively anything legal. Anthropic drew two red lines: no fully autonomous lethal weapons, and no mass domestic surveillance. It argued, among other things, that it had not validated those uses and did not trust them to work safely.

The talks broke down. On February 27, President Trump posted a directive ordering every federal agency to immediately stop using Anthropic’s tools. Shortly after, Defense Secretary Pete Hegseth’s department labeled Anthropic a “supply chain risk,” a designation historically reserved for foreign adversaries like Huawei and ZTE. That label does not just threaten a single contract. It requires defense contractors to certify they are not using Claude in their military work, which reaches deep into Anthropic’s broader business.

Anthropic sued. The company filed challenges in California federal court and in the federal appeals court in Washington, framing the blacklisting as unlawful retaliation for refusing to drop its safety restrictions. A judge temporarily blocked the designation while the case proceeds, and that litigation is still live. One judge reviewing the dispute reportedly described the government’s move in terms as stark as attempted corporate murder.

So when a fresh national security directive lands on Anthropic’s desk months later, citing a jailbreak the company calls trivial, it does not arrive in a vacuum. It arrives in the middle of an active, bitter, litigated fight over who controls how powerful AI gets used.

Why This One Is Different

Set the politics aside for a second and look at the precedent.

By multiple accounts, this is the first time a leading AI company has taken a publicly deployed model offline because the federal government told it to. Companies have delayed launches. They have restricted features. They have geofenced access by region. But pulling a shipped, in market frontier model, for every user, on a few hours notice, at government direction, is new territory.

The timing sharpens it further. Anthropic filed confidentially for an IPO earlier this month, reportedly disclosing a revenue run rate near $47 billion and a valuation around $965 billion. A sudden, government compelled blackout of the company’s flagship models is not the kind of stability story you want to be writing in the weeks before you go public.

And there is a chilling effect that reaches past Anthropic. If the lesson Silicon Valley takes from this is that disagreeing with the government on AI safety terms can get your products switched off by directive, that reshapes the calculus for every lab weighing how hard to push back. The balance of power between Washington and the AI industry just shifted, and everyone building at the frontier saw it happen.

What’s Still Unknown

A lot, honestly.

The directive itself did not spell out the specific national security concern. Anthropic’s understanding of the jailbreak is its own reconstruction, based on a demonstration and a report it believes underpins the order. The government has not publicly laid out its technical case. Anthropic has promised to share more detail over the next twenty four hours, and says it is working to restore access as fast as it can while insisting the whole thing is a misunderstanding.

Watch for three things in the next day. Whether the government articulates a concrete, specific harm rather than a general capability worry. Whether Anthropic’s promised technical disclosure holds up its trivial and ubiquitous framing. And whether this stays a one off, or becomes the template for how Washington handles frontier releases it does not like.

Because if it becomes the template, Friday night was not an incident. It was a preview.

That’s the game. Uncle AI. Operator Mode.