Ukraine's Ministry of Digital Transformation has drawn a blunt lesson from a three-week episode in Washington: when a foreign government can switch off your AI provider overnight, that provider is a liability in wartime. Roman Kyslyi, the ministry's Chief AI Officer, told Reuters on July 7, 2026, that Kyiv will now prioritize AI systems it can run on its own infrastructure over models that remain, by design, under a vendor's remote control — a category that includes Anthropic's and OpenAI's flagship offerings. "It confirms that AI sovereignty isn't just a defensive talking point, it's a necessity," Kyslyi said.
What actually happened
The trigger was concrete, not hypothetical. Anthropic publicly launched Claude Fable 5 and the more capable Claude Mythos 5 on June 9, 2026. Three days later, on the evening of June 12, the U.S. government ordered Anthropic to suspend access to both models entirely — not just for foreign nationals subject to export controls, but, in Anthropic's words, "for all our customers" worldwide, citing an unspecified national-security concern later traced to a narrow jailbreak that let the models assist with vulnerability discovery in code. Anthropic complied but pushed back hard, writing that it disagreed "that the finding of a narrow potential jailbreak should be cause for recalling a commercial model" already matched by competitors.
The suspension lasted 19 days. On June 30, Commerce Secretary Howard Lutnick notified Anthropic that a license was "no longer required" for the models' export or domestic transfer, after the company agreed to proactively flag security risks, coordinate release protocols with the government, and report malicious use. Fable 5 returned to global users on July 1. The whole cycle — launch, worldwide blackout, reversal — took three weeks, and it happened to a U.S. ally's most capable available AI tooling for reasons even Anthropic called a misunderstanding.
The case for Kyiv's move
That sequence is a legitimate reason for a wartime government to worry. Ukraine's digital ministry runs the Diia app, coordinates battlefield data pipelines, and increasingly leans on AI for translation, disinformation triage, and cyber-defense — functions that cannot tolerate a surprise 19-day outage triggered by a dispute Ukraine has no part in and no visibility into. The ministry's current workaround, routing Diia's assistant through Google's remote-hosted Gemini via EU servers, already requires stripping personal data before every query, precisely because, as Kyslyi put it, "we don't control those models." A regulator here would rightly note that dependence on infrastructure another government can unilaterally freeze is a single point of failure no serious continuity plan should accept, and Ukraine's plan to field its own model — a Gemma-based national LLM built with telecom operator Kyivstar, aimed for release across government, business, and military use by autumn 2026 — is a sensible parallel investment regardless of what Washington does next.
Where the policy is actually well-calibrated — and where the risk lies
What's notable is what Kyslyi did not say. He did not announce a ban on American AI, and he was explicit that the criterion is deployability, not vendor nationality: "If the vendor will provide it to run on our on-premise (infrastructure), there are no restrictions." That is a procurement standard, not decoupling — any provider, American or otherwise, that offers a self-hosted or sovereign-cloud deployment still qualifies. Framed that way, this is proportionate risk management, not techno-nationalism, and other governments watching the Fable/Mythos episode should take the same targeted approach rather than reaching for blanket exclusions.
The deeper risk sits on the U.S. side of this story. Frontier, provider-controlled models come with the heaviest safety tooling, the most active monitoring, and the fastest patching — precisely the properties that make them easier to govern responsibly. Open-weight alternatives like Gemma, which Ukraine is building its sovereign model on, are improving quickly but still trail frontier-tier systems on capability, and once weights are downloaded onto local servers, no export authority can recall them if something goes wrong. An export-control regime whose most visible recent act was an ambiguously justified, world-wide shutdown that its own author reversed inside three weeks does not project cautious stewardship — it projects unpredictability. And unpredictability is exactly what pushes allies toward less controllable, harder-to-monitor alternatives, which is a worse outcome for U.S. national security than the scenario the June 12 order was ostensibly designed to prevent.
The proportionate read
Ukraine's shift is not evidence that American AI companies are becoming unreliable partners — Anthropic's own account shows it fought the order and got it reversed within weeks. It is evidence that ad hoc, opaque export enforcement carries costs beyond the immediate dispute: it teaches capable, security-conscious allies to build around you rather than with you. The fix is not to relax legitimate national-security review of frontier models with genuine dual-use risk. It's for that review to come with the predictability, transparency, and speed that let allies keep planning their AI infrastructure around U.S. providers instead of away from them.