US order cuts Anthropic access

A critical AI system can disappear in the middle of an ordinary working day, even while the servers are still humming and the bills are paid. On 12 June 2026, Anthropic said the US government had ordered it to suspend foreign-national access to Fable 5 and Mythos 5, including access for its own foreign-national employees. For Europe, and for countries such as Ireland that now run banking, telecoms, research and public administration on foreign cloud infrastructure, the lesson was immediate: an export-control decision in Washington can become an operational outage at home.

The Safety Claim Came First

The order was understood to have followed a claimed bypass of Fable 5’s safety barriers. A bypass, or "jailbreak", means finding a way to phrase prompts so that a model ignores restrictions it is meant to follow. The building may be secure, the locks may work, but the person at reception is persuaded to hand over the wrong badge.

That distinction matters because these systems were sold as guarded rooms, not ordinary chatbots. Business Insider reported that Anthropic built Fable 5 to restrict requests in cybersecurity, biology and chemistry. Mythos was closer again to the sensitive side of the line: Le Monde had already placed it inside French cybersecurity discussions involving Orange and Crédit Agricole.

The public evidence did not match the scale of the consequence. Anthropic said in its own review that it understood the government had seen a claimed bypass, but judged the demonstration to involve minor vulnerabilities that were already known. AP reported that the Commerce Department gave no public timetable for restoring access and no detailed explanation. National-security officials may have reasons they cannot publish. Customers experienced the part that was visible: access stopped.

Nationality Became The Circuit Breaker

The legal machinery behind the order predates modern AI. Under US export rules, sharing controlled technology with a foreign national can count as an export even when that person is sitting in California and the data never leaves the cloud, as federal rules set out. Once that logic is applied to a live AI service, identity becomes part of the infrastructure.

That is a hard line to draw cleanly in cloud computing. Access moves through customer accounts, employees, contractors, API keys, regions and support desks. R Street argued that, without a reliable way to check citizenship at every doorway, the simplest compliant move is to shut access until the rule is clear. Reuters-linked reporting said Anthropic disabled access after the order and that AWS was asked to revoke access across regions.

For a bank, ministry or telecoms operator, the resilience question is not theoretical. If a legal order removes your main model tonight, which process is still working tomorrow morning?

France grasped the point quickly because it had actual users in view. Le Monde also tracked a broad political reaction, with figures from Jordan Bardella to Jean-Luc Mélenchon, Gabriel Attal and Édouard Philippe reading the move as evidence that dependence on AI had become a strategic exposure. Brussels chose more careful wording, but Euronews reported Commission spokesperson Thomas Regnier saying the EU was assessing the consequences for European users.

Europe Needs Backups, Not Slogans

The quickest European answer is not a slogan about sovereignty. It is architecture: avoid relying on a single model, keep fallback providers, require data and prompts to be portable, and decide which workflows are too sensitive to depend entirely on systems controlled abroad. Germany is already having that conversation. Tagesschau covered the incident as a resilience test for administrations and critical businesses, while still acknowledging that US AI remains part of procurement.

EU policy now lands in a sharper setting. The Commission presented its tech-sovereignty package on 3 June 2026, just days before the Anthropic shock. The proposed Cloud and AI Development Act says EU data-centre capacity should at least triple within five to seven years. InvestAI adds the industrial wager, with €200 billion for AI development and €20 billion for AI gigafactories.

Those plans may build leverage over time. They do not alter who controls a proprietary American model today. Europe can regulate AI, buy computing capacity and fund domestic champions, but critical systems also need the dull disciplines that keep services alive: exit routes, backups, contract clauses, open-source options where they are good enough, and public services that keep running when a foreign switch is pulled.