US government forces Anthropic to shut down its most powerful AI models over security fears

Anthropic disabled its two top-performing AI models on Friday, June 12 following an export control order from the U.S. Department of Commerce. It was an unusual instance of a government interfering with commercialization of frontier AI technology.

U.S. Commerce Secretary Howard Lutnick ordered Anthropic CEO Dario Amodei at 5:21 p.m. ET on Friday to restrict access to Claude Fable 5 and Claude Mythos 5, according to Anthropic’s statement.

The directive applied to all foreign individuals, regardless of whether they were inside or outside the United States, including Anthropic’s own foreign-born employees.

Anthropic stated that it couldn’t effectively enforce nationality-based restrictions and so deactivated the models. The other Claude models are still available.

Commerce Department moves against Anthropic’s newest models

The directive followed a warning from an unnamed company about a technique that could bypass parts of Fable 5’s safety system, Axios reported, citing an administration official.

Anthropic had released the two models only three days earlier, on June 9. Fable 5 was publicly available, while Mythos 5 was limited to selected cybersecurity partners under Project Glasswing.

Prior to this, the Trump administration had been engaged in conversations attempting to persuade the company to postpone its rollout but Anthropic proceeded regardless.

A White House official said they felt it important to keep the models locked down while the nation strengthens federal security infrastructure and said he thinks that will take weeks.

Under the directive, any party seeking to export, reexport, or transfer either model in the U.S. would need a license from the Commerce Department. Anthropic would also have to apply for individually validated licenses. Failure to comply could carry monetary and civil penalties, Axios reported.

Anthropic says the jailbreak risk is overstated

Anthropic pushed back against the directive, saying the jailbreak in question was narrow and did not amount to a universal bypass of the models’ safety systems.

The company described the technique as one that asks a model to review a specific codebase and identify software flaws. Anthropic said it reviewed a demonstration of the method and found that it surfaced only previously known, minor vulnerabilities.

“We have found that other publicly-available models are able to discover them as well without requiring a bypass,” Anthropic wrote, naming OpenAI’s GPT-5.5 as one example.

Anthropic also said it had conducted thousands of hours of red-teaming before launch with the U.S. government, the UK AI Safety Institute, and private organizations.

Those tests found no universal jailbreak, the company said, meaning no single method that could broadly defeat the model’s safety controls across many capabilities.

“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,” Anthropic wrote. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”

Fable 5 was already facing criticism

The shutdown capped a turbulent launch week for Anthropic.

Fable 5 debuted on June 9 with filters that routed sensitive queries in cybersecurity, biology, chemistry, and AI research to an earlier model, Claude Opus 4.8. Those filters were criticized almost immediately as too restrictive.

According to Cryptopolitan’s prior report, users said the model struggled with basic biology questions involving cell membranes and mRNA vaccines.

Another controversy centered on Anthropic’s 319-page system card, which said Fable 5 would automatically reduce output quality for questions related to frontier AI research without warning users.

Anthropic changed that policy on June 10 after researchers argued that hidden safeguards would concentrate AI capabilities among a small group of companies. The company apologized and said future safeguards would be visible to users.

Microsoft also restricted employee use of Fable 5 because of Anthropic’s 30-day data retention requirement for Mythos-class models, The Verge reported.

Under that policy, prompts and outputs are stored for 30 days for trust and safety purposes, while flagged content can be retained for up to two years.

Anthropic says it is working to reverse the shutdown

Anthropic said the government’s action was based on a misunderstanding and that it is working with officials to resolve the issue. The company said it would provide more information within 24 hours.

The shutdown comes at a critical moment for Anthropic. The company submitted confidential documents earlier this month for a planned U.S. IPO after raising $65 billion at a $965 billion valuation late last month, Reuters reported.

Anthropic is now facing pressure on two fronts: regulators are questioning the safety of its most advanced models, while investors are betting on its ability to keep growing commercial revenue.

Claude’s official account on X told users that new sessions will now default to their selected model or to Opus 4.8. Existing Fable 5 sessions will end with an error, and developers were advised to update their integrations to other Claude models.