Europe’s AI sovereignty just became a security emergency

In March 2025, NATO acquired the Maven Smart System, an AI-powered targeting platform built by the American defence contractor Palantir that runs on AI models developed by US companies. One year later, that same system helped coordinate the bombing of over 1,000 targets in Iran in 24 hours.

No European parliament debated this acquisition. No European government voted on the terms under which its military would depend on American AI for battlefield decisions. And the only company that tried to maintain ethical constraints on how its AI could be used in warfare was blacklisted within 72 hours. 

This is no longer a technology policy debate. It is a European security emergency. 

The facts are straightforward. Anthropic, the company that built the AI model Claude, the first cleared for use on the Pentagon’s classified networks, had set two conditions: its technology could not be used for mass surveillance of American citizens and could not power fully autonomous weapons. The Pentagon demanded these restrictions be removed. When Anthropic refused, it was designated a “supply chain risk to national security”. A label previously reserved for foreign adversaries like Huawei,  and replaced by competitors offering no such constraints. OpenAI signed a deal with the Pentagon the same evening. 

The market signal was immediate and unambiguous: ethical constraints on military AI are a business liability. Any company that maintains them will be eliminated and replaced by one that does not. 

Europeans should understand what this means for them. The Maven Smart System now deployed across NATO operates on the same American AI infrastructure, under the same “all lawful use” doctrine that the Pentagon has imposed on its AI providers. Twenty thousand US military personnel already use it daily across every major combatant command. It integrates dozens of data sources – such as satellites, surveillance, signals intelligence – into a single cognitive architecture through which targeting decisions flow. 

During the Iran operation, this system generated targets at a rate of one every 86 seconds. One artillery unit, using AI-powered targeting, performed the work of 2,000 staff with a team of 20. The Pentagon insists human operators took all final decisions. But what does a “decision” mean at 86-second intervals? To genuinely evaluate each target, an officer would need to independently verify intelligence, assess collateral damage, and consider proportionality under international humanitarian law. At that tempo, the human in the loop can only approve what the algorithm has already decided. Oversight becomes a procedural formality, not a substantive act of judgment. 

This is the system NATO has bought. It is, however, not the hardware; it is the cognitive infrastructure. The architecture through which military thinking itself now occurs. 

Europe has spent years building the AI Act, the most comprehensive AI regulatory framework in the world. But the AI Act regulates civilian high-risk systems. It says nothing about cumulative military AI dependency. It does not address the question of what happens when the decision-making infrastructure of European defence is owned, operated, and constrained (or unconstrained) by a handful of American companies answering to the Pentagon’s terms. 

The Anthropic case demonstrated three issues that European policymakers must consider. First, corporate self-regulation of military AI is dead. The one company that tried it was punished so swiftly and so publicly that no rational actor will attempt it again. Second, the pace at which AI-assisted targeting operates makes traditional concepts of human oversight meaningless at scale. International humanitarian law, built on individual commander responsibility, has no answer for algorithmic targeting at 86-second cycles. Third, and most critically for Europe, strategic autonomy is impossible when the cognitive infrastructure of your military depends on technology whose terms of use are dictated by a foreign government.

None of this requires Europe to reject AI in defence. That option does not exist. But it does require three urgent steps. NATO members must establish binding minimum decision times for AI-assisted targeting, not procedural checkboxes, but enforceable standards that make genuine human judgment possible. The EU must extend its AI governance framework to address military AI dependency, including transparency requirements for the AI systems integrated into allied command structures. And Europe must invest in sovereign AI capability for defence. Not to compete with American models, but to ensure that the infrastructure through which European soldiers make life-and-death decisions is not subject to terms imposed unilaterally by Washington. 

The window for these decisions is narrow. Every military in the world watched the Iran operation. Every general saw that AI-assisted targeting works at scale. The incentive to replicate, and to do so using whatever American AI infrastructure is available, on whatever terms are offered, is overwhelming. If Europe does not act now to define the conditions under which it will use military AI, those conditions will be defined for it. 

Vladimír Šucha is a visiting professor at the European University Institute in Florence and founder of resilientcore.ai. He served as Director General of the European Commission’s in-house Scientific Service (2014–2019) and as Head of the European Commission’s Representation in Slovakia (2021–2024).