In mid-June 2026, the US government took an unprecedented step in the country's AI policy: Anthropic was ordered to deactivate its two recently launched models, Fable 6 and Mythos 5, for all users. According to the research source, the decision stems from fears that the models' safety mechanisms could be circumvented — and that in the wrong hands, they could write functional code for cyberweapons and identify critical vulnerabilities in national infrastructure.

Government fears cyber threat

The rationale behind the shutdown order is specific: authorities believe the models, if "jailbroken," possess an "unparalleled ability to generate advanced exploits and locate critical flaws in infrastructure," according to the research source. The concern is therefore not about normal commercial use, but about scenarios in which safety barriers are deliberately dismantled.

The shutdown affects more than just external customers — according to the source material, Anthropic was also required to revoke access for its own employees holding foreign citizenship, in order to comply with the order. Dean Ball, a former White House official, told TechCrunch that this could mean users will in future be required to document their citizenship to gain access to Anthropic's models.

"We disagree that the discovery of a narrow potential jailbreak should be grounds for withdrawing a commercial model used by hundreds of millions of people" — Anthropic
US shuts down Anthropic's most powerful AI — company fires back - Bilde 1

Anthropic: Core principles have been violated

The company makes no effort to conceal its dissatisfaction. In a public blog post, Anthropic states that it considers the jailbreak issue "narrow and non-universal," and that comparable capabilities exist in other freely available models. The company accepts in principle that authorities should have the power to halt unsafe AI systems, but stresses that any such process must be "transparent, fair, clear, and grounded in technical facts" — criteria it believes this decision fails to meet, according to TechCrunch.

A historic step in AI regulation

This marks a significant shift in how the United States manages AI risk. Previously, American measures have primarily focused on export controls for processor chips and development tools — not the direct shutdown of commercial AI models aimed at a global audience.

For the first time, a Western government has ordered the withdrawal of a commercially launched AI model on security grounds

A broader regulatory debate is simultaneously unfolding in the US. The recently reintroduced FAIR Elections Act seeks to ban AI-generated electoral content intended for voter suppression. The Sectoral AI Governance Act of 2026 would give federal agencies clearer authority to enforce rules where AI is used to circumvent existing legislation, according to the research source. Anthropic's case throws these discussions into sharp relief: what happens when a company's own safety warnings about AI risk are used as justification for government intervention?

The irony of the whistleblower's role

There is a considerable irony in Anthropic's situation. The company has consistently positioned itself as one of the most responsible actors in the AI industry, placing strong emphasis on security risks in its public communications. It is precisely this openness about potential dangers that may now have helped trigger the government's response — a response the company itself considers misguided.

The case raises fundamental questions about the relationship between AI companies' responsibility to flag risks and the state's authority to act on such warnings. How this conflict is resolved could set a precedent for AI regulation well beyond US borders.