In June 2026, the artificial intelligence company Anthropic found itself at the centre of an unprecedented clash between commercial AI development and national security policy. The United States government issued an export control directive ordering the company to suspend access to its newest and most capable models, Claude Fable 5 and Claude Mythos 5, for all foreign nationals. Because the order applied to foreigners both inside and outside the United States, including Anthropic’s own employees, the company concluded that selective compliance was technically impossible and disabled the two models worldwide. It was a striking moment: software used by hundreds of millions of people was switched off, not by a market decision, but by a government order.
According to Anthropic, the directive arrived at 5:21pm on a Friday and did not spell out the government’s specific security concern in detail. The company said it had only received partial information, but that after reviewing a report it believed had triggered the order, its experts concluded the issue centred on a relatively narrow capability: the models’ aptitude for reviewing program code and identifying software vulnerabilities. Mythos 5 in particular is regarded as exceptionally skilled at detecting flaws, some of which had reportedly gone undiscovered for decades, a strength that US authorities and selected companies had used to harden their own systems.
The government’s underlying worry, as reported across multiple outlets, was that such capabilities could be turned into a cyberweapon if placed in the wrong hands. Concerns reportedly included the possibility of bypassing, or “jailbreaking,” the safeguards intended to limit the use of Fable 5 for offensive cybersecurity tasks. Some reporting also linked the timing of the order to suspicions that a China-linked group had gained access to Anthropic’s new model, sharpening fears that frontier AI could accelerate the discovery of exploitable weaknesses in critical infrastructure.
Anthropic publicly disputed the rationale. The company argued that the capability in question is not unique to its products, noting that rival systems such as OpenAI’s GPT-5.5 possess similar code-analysis abilities, and that the safety measures built into Fable 5 had been extensively tested. It maintained that blocking software relied upon by a vast global user base was a disproportionate response to a narrow concern, and warned that the standard implied by the directive could, if applied consistently, halt the deployment of virtually every new frontier model across the industry.
The episode did not emerge in a vacuum. Earlier in 2026, tensions between Anthropic and Washington had already escalated dramatically. Following a dispute over the Pentagon’s preferred contract terms, which would have permitted purchased AI models to be used “for any lawful purpose,” Anthropic sought carve-outs to prevent its models from being deployed in autonomous weapons systems or mass domestic surveillance. The disagreement prompted a federal directive for agencies to step back from the company’s products, and in early March the Pentagon designated Anthropic a “supply chain risk,” instructing the military to stop using its models and barring defence contractors from relying on them for government work.
The fallout extended well beyond the Department of Defense. A range of civilian agencies and institutions that had piloted or integrated Claude under the General Services Administration’s “OneGov” arrangement, including bodies in health, space research, and the national laboratory system, were affected when that agreement was pulled. Policy analysts noted that a dispute originating with a single defence procurement decision had rippled outward, raising difficult questions for civilian agencies at every level of government about the durability of their AI deployments.
Anthropic has not accepted these measures quietly. The company launched legal challenges in two federal venues, the Northern District of California and the D.C. Circuit Court of Appeals, contesting the supply chain risk designation under separate statutes. A federal judge granted a preliminary injunction restraining the Pentagon and other agencies from blocking the use of Claude, while the Department of Justice signalled its intention to appeal, ensuring the legal contest will continue. Throughout, Anthropic has stressed that its other models, including Claude Opus 4.8, remain available to users.
For businesses, developers, and policymakers, the standoff offers an early glimpse of how the next phase of AI governance may unfold. It illustrates how export controls, originally designed for physical goods and chips, are being stretched to cover software that can be disabled remotely and instantly, and how national security imperatives can collide with the realities of globally distributed products and workforces. Whatever the courts ultimately decide, the message for organisations building their operations around frontier AI is unambiguous: model availability is no longer a purely technical or commercial matter, but one increasingly shaped by geopolitics, and resilience now means planning for the possibility that a critical tool could be switched off overnight.
