European Governance Lab

Menu
Sign in
Get Started

Shopping Cart

No products in the cart.

Shopping Cart

No products in the cart.

Close search

Governing the National Security Exception in EU AI Law

  • Digital Policy & Technology, EU Institutions & Governance, Rule of Law & Justice
  • June 26, 2026

Introduction

While the strategic significance of artificial intelligence is widely acknowledged, the primary legal issue for the European Union concerns whether the national security exception is evolving into a structural governance mechanism that enables strategically significant AI systems to circumvent the Union’s principal accountability framework.

The EU AI Act entered into force on 1 August 2024 and will apply progressively over the coming years (subject to any revisions in the meantime). Article 2(3) excludes AI systems developed or used exclusively for military, defense, or national security purposes, while Recital 24 clarifies that AI systems used for both civilian and military, defense, or national security purposes remain within the scope of the Regulation insofar as they are deployed for civilian, humanitarian, law-enforcement, or public-security functions. The exclusion is therefore functional rather than actor-based: the same system may fall outside the Regulation in one exclusively national-security context while remaining subject to EU rules in another civilian or dual-use setting. [1](European Union, 2024) 

The same competence-preserving logic appears elsewhere in EU digital law. The Data Act provides that it does not affect Member State competences concerning public security, defence, or national security; whereas the NIS2 Directive excludes public administration entities whose activities are predominantly carried out in the areas of national security, public security, defence, or law enforcement, and permits Member States to exempt certain entities carrying out such activities from specific obligations. At the same time, the case law of the Court of Justice suggests that national-security rationales do not automatically displace EU legal standards where measures engage sectors, actors, or activities otherwise governed by Union law. [2](European Commission, 2024; European Union, 2023, 2022; CoE, 2024)

The EU exception as a constitutional blind spot

The legal obstacle is obvious. Article 4(2) TEU states that national security remains the sole responsibility of each Member State. Yet that clause does not create a zone of complete legal immunity. [3](European Union, 2025) 

In Privacy International and La Quadrature du Net, the Court of Justice of the European Union held that national-security measures do not fall outside EU law merely because they pursue national-security objectives, where they impose processing, retention, access or transmission obligations on actors governed by EU law, such as electronic communications providers. The Court further held that general and indiscriminate interferences with communications data remain subject to review under Directive 2002/58, EU law and the Charter, including strict necessity, proportionality, time limitation, safeguards against abuse, and effective independent or judicial review.   National security is thus constitutionally weighty, but not wholly beyond legal discipline. [4](CJEU, 2020a; CJEU, 2020b)

This distinction is particularly significant in the context of artificial intelligence. The AI Act’s exclusion applies regardless of the type of entity carrying out the activity, including private vendors, where AI systems are placed on the market, put into service, or used exclusively for military, defence, or national security purposes.  However, neither the AI Act nor EU primary law provides a precise or operational definition of “national security,” leaving the boundary between national security, public security and law enforcement legally fluid and context dependent. The boundary with public security remains ambiguous, especially in areas where surveillance, migration control, cybersecurity, and law enforcement intersect.

Secondary analysis has therefore identified three acute EU vulnerabilities: relabelling civilian or dual-use systems as security tools; outsourcing sensitive functions to private suppliers while avoiding AI Act duties; and fragmentation across Member States because domestic understandings of national security differ. These are precisely the conditions in which exceptions tend to expand. [5](Powell, 2024; Vogiatzoglou, 2024)

Comparative context

In the United States, federal civilian AI governance memoranda expressly do not apply to AI used as part of a “National Security System”, while defense AI is governed separately through the Department of Defense’s responsible-AI policy and Directive 3000.09 on autonomy in weapon systems. [6](U.S. Department of Defense, 2022, 2023)

The 2025 AI Action Plan further places economic competitiveness, industrial mobilization, and national security at the center of national AI policy, emphasizing technological leadership, defense-industrial expansion, cyber resilience, and strategic competition with China. [7](OMB, 2024; White House, 2025) This broader securitized approach is reinforced in the 2026 U.S. National Defense Strategy, which explicitly integrates artificial intelligence into military modernization, deterrence strategy, cyber defense, and defense-industrial mobilization, while framing technological superiority as essential to maintaining U.S. strategic primacy. The impact of any purge on previous administration’s ‘restrictive rules’ remains to be seen. [8](U.S. Department of War, 2026)

The outcome is not a completely unregulated environment, but rather a strategically integrated executive-security framework that operates largely outside conventional civilian AI governance structures.

The United Kingdom follows a similar separation through a different model. Its “pro-innovation” framework relies on existing civilian regulators rather than a comprehensive AI statute, while defense AI is handled through the Ministry of Defense’s own strategy, JSP 936, and an AI Ethics Advisory Panel. [9](UK Government, 2023, 2024)

By contrast, China does not carve national security out of AI governance so much as embed AI governance within a wider security state: the 2023 generative-AI measures were adopted to promote development while safeguarding national security and social public interests, and the 2025 global-governance action plan and related data-security rules continue that securitized approach. [10](Government of China, 2023, 2025a, 2025b)

In this comparative context, the European Union faces a unique risk: not the executive separation characteristic of the United States, nor the security integration model of China, but rather the insufficiently governed expansion of exceptions within a constitutional framework otherwise dedicated to rights-based oversight.

A calibrated framework compatible with Article 4(2) TEU

A workable EU response is therefore procedural, not substantive. The Union should not attempt to decide whether a Member State faces a genuine national-security threat; Article 4(2) TEU makes that constitutionally fragile. It can, however, require disciplined procedures whenever a Member State invokes AI Act Article 2(3) for an AI system procured on the internal market, supplied by a private provider, or likely to generate spillovers into law enforcement, migration, public security, or fundamental-rights protection. [11](European Union, 2024)

The first mechanism should be a confidential National Security AI Justification created before deployment. That document should record: the precise security purpose; why the case is national rather than merely public security; whether the system is dual-use; the data sources used; the provider and supply-chain dependencies; the expected rights impacts; human-oversight safeguards; cybersecurity and robustness testing; and why less intrusive means are insufficient. It should also contain a mandatory sunset clause, for example, 12 months, after which the exemption lapses unless renewed on new grounds. This would operationalize proportionality rather than merely invoke it. [12](Vogiatzoglou, 2024)

The second mechanism should be layered oversight. Each Member State should designate an independent review body (judicial, parliamentary-intelligence, or a specialized security commissioner) with authority to inspect the justification, demand mitigation, and terminate reliance on the exception when the criteria are not met. Ex post judicial review should remain available, at least in confidential form, where secrecy is unavoidable. At the EU level, Member States should transmit a restricted metadata notification to the European AI Office and the AI Board: not operational intelligence, but information on the system category, provider, legal basis, duration, renewal, and aggregate rights incidents. 

This is institutionally plausible because the AI Act already contains procedural templates that protect secrecy without abandoning accountability: fundamental rights impact assessments for certain high-risk uses, post-market monitoring obligations, and a secure, non-public section of the EU database for law enforcement and migration AI. The AI Office already exists to support the coherent application of the AI Act and coordination among national authorities. A confidential notification channel would therefore respect Article 4(2) TEU while allowing the Union to detect patterns of misuse, issue guidance on the national/public-security boundary, and reduce fragmentation across the internal market. [13](European Union, 2024; European Commission, 2026). Such coordination would not empower Union institutions to approve, reject, or reassess Member State national-security determinations, but rather to supervise procedural consistency, proportionality safeguards, and systemic spillover effects within the internal market.

Conclusion

The national security exception should not be treated as a legal vacuum, but as a domain requiring tailored governance mechanisms compatible with constitutional limits. Article 4(2) TEU preserves Member State responsibility for national security, yet the growing strategic importance of artificial intelligence increases the risk that broadly framed security claims may gradually displace the accountability structures underpinning the Union’s digital constitutional order.

The central issue is therefore not whether Member States retain authority over national security AI, but whether the invocation of that authority remains legally traceable, procedurally disciplined, and subject to review. This becomes particularly significant where private vendors supply systems later classified as exclusively national-security tools, or where dual-use technologies move across defense, intelligence, migration, law-enforcement, and public-security contexts.

A calibrated framework based on confidential justification, sunset clauses, independent oversight, and restricted EU-level notification would preserve operational secrecy while limiting opaque and fragmented exception practice. Such a framework would not require the Union to reassess national-security threats, but would ensure that recourse to Article 2(3) remains subject to legality, necessity, proportionality, and procedural accountability. Ultimately, the future of the AI Act’s national security exception is not merely a question of competence allocation, but of whether exceptional AI governance can remain compatible with the rule-of-law principles on which the Union is founded.







References

Have research worth publishing?

Share your research with EGL's community of EU professionals and aspiring policymakers

Join Now

Ioannis Sidiropoulos

Lawyer, LL.M (LSE, UvA), MA (UNIC), PhD (c)

Lawyer, LL.M (LSE, UvA), MA (UNIC), PhD (c) in AI Sovereignty and National Security (EUC)

Back to Publications

Report

Harassment or bullying behavior
Contains mature or sensitive content
Contains misleading or false information
Contains abusive or derogatory content
Contains spam, fake content or potential malware

Block Member?

Please confirm you want to block this member.

You will no longer be able to:

  • Mention this member in posts
  • Invite this member to groups
  • Message this member
  • Add this member as a connection

Please note: This action will also remove this member from your connections and send a report to the site admin. Please allow a few minutes for this process to complete.

Report

You have already reported this .