The US government ordered a major AI model offline — and set the price of its return

For the first time, the United States government used an export control — a rule restricting who outside the country can receive an American product or technology — to force an artificial intelligence model offline. And the company that makes it agreed to a new layer of government oversight to switch it back on. On June 12, 2026, the Commerce Department directed Anthropic to cut off access to Claude Fable 5 and Claude Mythos 5, its newest AI models, just three days after their launch. The models went dark for users everywhere. Fable 5 came back online July 1, after Anthropic agreed to give the government early access to future models, share intelligence about attacks, and add a new safety filter.

The episode set two precedents at once. It showed that Washington can, in practice, turn off an AI tool used by a vast number of people overnight. And it established — through a private negotiation rather than any law — what an AI company must give up to get such a tool turned back on.

Three words started it

The trouble began, according to reporting by NBC News and Axios based on unnamed sources, with researchers at Amazon testing Fable 5 before its release. By prompting the model with the phrase “fix this code,” they found it would identify software vulnerabilities — flaws in computer code that attackers could use to break in — and, in one case, write code demonstrating how to exploit one.

Amazon flagged the finding to the Trump administration. Accounts differ on exactly who Amazon CEO Andy Jassy called first, but within days the government acted — turning an export control against access to an AI model, something it had never done before.

The government’s order has not been made public. Neither has the Amazon report that triggered it. Everything known about both documents comes secondhand, from Anthropic and from unnamed officials.

Why the whole world lost access

According to Anthropic, the directive arrived at 5:21pm Eastern time on June 12 and barred access for “any foreign national, whether inside or outside the United States” — meaning anyone who is not a US citizen or permanent resident, wherever they live. Anthropic had no way to check users’ nationality in real time. So complying meant switching the models off for everyone, everywhere.

Anthropic protested immediately. “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” the company said, using the industry term for a trick that gets an AI model past its safety rules. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers” — “frontier model” being the industry’s term for the most advanced AI systems available at a given time.

Legal scholars writing at Just Security and Lawfare questioned whether the order was even lawful, noting that export controls traditionally govern products and technology crossing borders, not online access to software hosted in the US. At least one AI company has sued, according to the trade publication Export Compliance Daily, arguing the Commerce Department exceeded its legal authority. That question remains open.

Was it even dangerous?

Much of the security community sided with Anthropic. On June 15, an open letter organized by Alex Stamos, the former Facebook security chief, and signed by nearly 80 security leaders at the time, per Axios (the count grew afterward), argued that Fable 5 is “not uniquely good” at finding software flaws — other AI models can do the same — and that taking it away hurt the defenders who use it to find bugs before criminals do.

Katie Moussouris, a well-known security researcher who reviewed the Amazon report at Anthropic’s request, went further. Finding vulnerabilities, she said per Fortune, “is not a guardrail bypass” — guardrails being the safety limits built into a model — and called it “the most valuable thing an AI model can do for defensive security.”

The White House saw it differently. AI adviser David Sacks wrote on X, the social media platform, that “a highly credible trusted partner” of both Anthropic and the government had “come forward with a jailbreak of those guardrails,” that the administration had asked Anthropic CEO Dario Amodei to fix the problem or pull the model, and that “Dario refused” — a characterization Anthropic’s public account disputes. The core question of whether the behavior was a dangerous flaw or a useful feature was never formally settled. It was resolved by negotiation.

The price of turning it back on

That negotiation produced five commitments, described in Anthropic’s June 30 announcement. The company deployed a new safety classifier — an automated system that screens requests and blocks dangerous-looking ones — which Anthropic claims stops the reported technique more than 99% of the time. No outside party has published a test of that figure. Anthropic also acknowledged the filter wrongly blocks some harmless coding requests, and routes blocked users to an older model.

Beyond the filter, Anthropic agreed to give the government access to future frontier models before public release, to rapidly share information about jailbreaks and threats, to dedicate staff and computing power to joint government research, and to build a framework with Amazon, Microsoft, and Google for scoring how severe a jailbreak is.

“Over the past two weeks, we have worked closely with Anthropic to analyze and approve Fable 5…” Commerce Secretary Howard Lutnick wrote on June 30. Mythos 5, the version with fewer safeguards available only to vetted organizations, had already been restored to roughly 100 vetted US organizations on June 26, according to Anthropic and Nextgov. Fable 5 returned globally on July 1.

What this means for you

If you use Claude, Fable 5 is back as of July 1. Anthropic limited paid users to half their usual weekly usage through July 7 while it scaled capacity back up, per its redeployment announcement. Anthropic’s support documents also say Fable 5 conversations are now retained for 30 days so the company can research jailbreak attempts — worth knowing if you care about data privacy.

If you use AI for programming, expect occasional wrong blocks: Anthropic admits its new filter flags some harmless coding work, and it is not yet clear whether the model is now less useful for legitimate security research.

The bigger lesson applies to everyone: an AI tool your work depends on can be switched off by government order with no warning and no published rule. And new models may arrive more slowly — OpenAI is reportedly limiting its new GPT-5.6 to select partners while it negotiates its own release terms with the government, according to Nextgov.

Whether pre-release government review becomes the permanent price of shipping powerful AI is the question to watch. No law currently defines when Washington can order a model offline, the reported lawsuit challenging the government’s authority is pending, and even Anthropic says the rules it just accepted “should be codified in strong regulation and applied equally across frontier model developers.” Until that happens, the switch stays where June proved it to be: in the government’s hands.

Disclosure: Plainly’s editorial tools are built on Claude models made by Anthropic, the company in this story.