US Blocks Anthropic's AI After 'Jailbreak' Claim—3-Day-Old Model Offline

For cybersecurity professionals, the abrupt suspension of Anthropic's Fable 5 and Mythos 5 over a suspected jailbreak is a case study in AI vulnerability management. The U.S. government acted on verbal evidence, while Anthropic's own review found only known, minor bugs—raising questions about how model providers should assess and disclose bypass risks, and whether export controls can effectively prevent the proliferation of dual-use AI capabilities. The incident underscores the fragility of even safety-oriented models.

On June 12, 2026, at 5:21 p.m. Eastern Time, the U.S. government issued an export control directive to Anthropic, ordering the artificial intelligence company to block access to its two most capable AI models—Fable 5 and Mythos 5—by any foreign national. Citing “national security” authorities, the order covered foreign nationals both inside and outside the United States, including Anthropic's own foreign-national employees. In response, Anthropic suspended access to both models for all users worldwide, the first such broad denial of service triggered by a government directive on a commercially available AI product.

“

For cybersecurity professionals, the abrupt suspension of Anthropic's Fable 5 and Mythos 5 over a suspected jailbreak is a case study in AI vulnerability management.

The directive landed just three days after Anthropic began rolling out Fable 5 on June 9, offering it free to all Pro, Max, and Enterprise customers through June 22. The company described Fable 5 as the safeguarded sibling of Mythos 5; both share the same underlying powerful model, but Fable adds guardrails that block or divert sensitive cybersecurity, biology, and chemistry queries. Mythos 5, unrestricted, was available only to vetted government cyberdefense and life‑sciences partners. The abrupt suspension therefore not only wiped out public access to a model millions had already used, but also disrupted critical national security and research programs.

Anthropic disclosed that the government provided only “verbal evidence of a potential narrow, non‑universal” jailbreak—a method to bypass the model’s safety mechanisms. After reviewing the demonstration, Anthropic concluded it revealed only “minor, already‑known bugs,” the sort that other publicly available models could identify without any exploit. This gap between the government’s threat assessment and Anthropic’s technical review is central to the legal and policy debate that now unfolds.

The order’s extraterritorial reach is extraordinary. It bars foreign‑national access anywhere in the world, forcing Anthropic to disable the models entirely to comply. This approach mirrors the scope of the International Traffic in Arms Regulations, but applied to intangible AI services accessible through APIs. It immediately confronts AI providers with a compliance nightmare: how to verify user nationality in real time, especially when models are served globally. For multinational companies and research collaborations, the order signals that any cloud‑hosted advanced AI could be cut off overnight if the U.S. government perceives a national security risk, however poorly defined.

The incident has already caused diplomatic ripples. Kanishka Narayan, the UK’s Minister for AI and Online Safety, publicly noted that the suspension affected UK customers and framed it as a case for technological sovereignty, explicitly linking the event to the UK government’s £1.1 billion investment in domestic AI chips. This narrative is likely to be echoed by other nations, accelerating a decoupling of the global AI ecosystem. If foreign enterprises and governments cannot rely on steady access to U.S.-based AI services, they will invest more aggressively in homegrown alternatives, potentially fragmenting AI safety standards and research cooperation.

For the AI industry, the long‑term implications are profound. First, trust in U.S. AI providers is damaged; a model marketed as free and open can disappear within hours on verbal government allegations. Second, the order sets a low evidentiary bar for export controls on AI, which could lead to precautionary self‑censorship by companies that fear running afoul of vague national security standards. Third, the episode may chill the development of safety research: if building robust safeguards into a model makes it a target for export rules, companies will face a perverse incentive to keep advanced capabilities restricted from the start, undermining the very safety goals the order supposedly protects.