Four days after Anthropic released Claude Fable 5 and restricted Mythos 5 to selected partners, a US export-control directive forced access to both models offline for foreign nationals. Because nationality could not be filtered in real time, Anthropic briefly disabled service broadly while it worked to comply.
The incident transformed AI sovereignty from a policy slogan into an operational question. Customers outside the United States learned that access to a deeply embedded model could disappear through a government letter, even when their contracts, data and workloads remained unchanged.
A safety dispute created a global service interruption
Washington cited a jailbreak affecting Fable 5 and Mythos 5. Anthropic characterised it as a limited ability to review code and identify errors, while government figures argued that the vulnerability was serious and should be fixed before deployment.
The disagreement arrived amid an already hostile relationship. Anthropic had opposed use of its technology for mass surveillance and fully autonomous weapons, the US government had labelled it a supply-chain risk, and the company was challenging that designation in court.
Regardless of the politics, the operational mechanism was stark. Export, re-export or domestic transfer required approval, and foreign-national access became the controlling factor. Employees and customers could lose a tool because a provider lacked identity controls granular enough to enforce the order selectively.
This makes jurisdiction part of reliability engineering. A model can remain technically healthy and commercially paid for while becoming legally unavailable between one task and the next.
Europe and Canada saw dependency rather than one jailbreak
European officials and politicians responded by questioning reliance on technology designed and controlled elsewhere. Canada’s prime minister similarly framed the episode as evidence that dependence on a small group of American providers creates strategic exposure.
The timing strengthened that reaction: Europe had just published a technological-sovereignty package covering cloud and AI development. The shutdown supplied a vivid example of why local compute, models and legal control may matter.
Sovereignty does not require every country to reproduce the largest frontier system. It can mean retaining critical data locally, supporting regional models for essential workloads and maintaining the technical ability to route tasks when one provider becomes unavailable.
Enterprises should identify which processes would stop if a model disappeared for a day, a month or permanently. A conversational pilot may tolerate delay; cyber defence, research or public services may require tested alternatives.
Resilience requires portability beyond a procurement promise
Multi-model architecture helps only when applications are genuinely portable. Proprietary tool calls, prompts, safety filters and evaluation thresholds can make a nominal backup unusable without extensive rework.
Teams need abstraction at the model interface, portable retrieval data and evaluation suites that show whether an alternative meets the task’s minimum quality. They should rehearse the switch before a crisis and understand how cost, latency and governance change.
Contracts can address notice, data export and transition support, but no supplier can promise immunity from its home government. Critical deployments should document jurisdiction alongside uptime and security.
Identity design also deserves attention. If regulations distinguish employees or users by nationality, providers may face pressure to collect more sensitive information. Organisations must weigh compliance against privacy and discrimination risks rather than adopt crude controls in haste.
The immediate dispute may end with a patch and licences. The strategic lesson will remain. Frontier AI is increasingly infrastructure, yet access is governed by corporations and states whose priorities can change abruptly.
Sovereignty is not simply owning a national model. It is knowing who can switch off a capability, understanding the consequence and having a credible way to continue when the answer lies outside the organisation.