US bans Anthropic's 2 top AI models after jailbreak claim, reshaping governance

For AI researchers and engineers, the abrupt recall of two leading AI models over a vaguely described jailbreak represents a watershed: national security is now directly shaping which models can exist and for whom. The case tests the robustness of model safety mechanisms and raises hard questions about transparency, evaluation standards, and the future of cross-border AI collaboration.

On June 12, 2026, Anthropic announced it will abruptly disable its most advanced artificial intelligence models—Fable 5 and Mythos 5—for all users, after the U.S. government ordered it to suspend access to those models for all foreign nationals. The export control directive, received at around 5:21 p.m. ET, cited national security concerns related to a potential “jailbreak” method that could bypass a safeguard, allowing the model to identify software vulnerabilities. This action represents a pivotal moment in AI governance: for the first time, a frontier AI model has been effectively pulled from the global market not by litigation or regulation, but through export controls traditionally used for physical goods like semiconductors. The decision immediately affects hundreds of millions of users and escalates an already fraught relationship between the Trump administration and the IPO-bound company.

“

On June 12, 2026, Anthropic announced it will abruptly disable its most advanced artificial intelligence models—Fable 5 and Mythos 5—for all users, after the U.S.

The background to this clash is essential. Earlier in 2026, Anthropic refused to permit the U.S. military to use its AI models for domestic surveillance and fully autonomous weapons systems. In response, the Pentagon designated Anthropic as a supply chain risk—effectively a blacklisting—and the company sued the Defense Department. Lawsuits over that blacklisting remain pending. While some tensions were reportedly easing across parts of the government, this new directive shows that the core conflict remains unresolved and is now spilling into export control territory. Anthropic is complying, stating that the “net effect” of a foreign-national-only restriction would be impossible to enforce without a full global disable.

The government’s justification is striking for its opacity. Anthropic says it was given only “verbal evidence of a potential narrow, non-universal jailbreak,” with no written documentation of the national security threat. The company disputes the severity, arguing that a narrow, known vulnerability, which could also be identified by other publicly available models, does not warrant recalling a commercial product that had been safely deployed. This dispute highlights a growing tension over how to evaluate jailbreak risks: regulators are apparently willing to act on classified intelligence with minimal public explanation, while developers demand more transparent, reproducible threat assessments.

The shift from controlling the hardware that powers AI to directly restricting access to AI models is a major escalation. For years, U.S. export controls concentrated on advanced chips and manufacturing tools, most notably in the sweeping October 2022 rules against China. Now, an AI model itself is treated as a controlled item under national security export regulations. This could set a precedent that any AI model deemed risky—based on classified information—could be removed from global circulation by executive fiat. For the AI industry, which relies on global developer communities and cross-border research, the chill could be immense. Companies may be forced to build geo-fenced silos or redirect R&D away from frontier capabilities.

For Anthropic, the immediate business impact is significant. The company is preparing for an IPO and counts its Fable and Mythos lines among its most advanced offerings, competing with models from OpenAI and Google. Disabling them for an extended period could erode user trust, delay revenue growth, and depress valuation just as public markets are scrutinizing AI business models. While other Anthropic models remain unaffected, the perception of vulnerability to sudden regulatory action may weigh on investor sentiment. Competitors that can demonstrate government alignment or that operate in less regulated jurisdictions may capture displaced users.