How can the European Union identify AI threats before they become security risks?

AI developments in China and Russia are increasingly becoming a security concern for the European Union. Yet current approaches to assessing these risks remain fragmented and reactive. They often treat “AI” as a single technology and developments in autocratic states as a single threat category. This can make it difficult to identify where the most serious security risks are actually emerging and to distinguish between different pathways to risk.

In a new AI4POL Policy Brief, Jens Prüfer, Kun He and Alexander Kriebitz examine how Europe can better anticipate emerging AI-related security threats and develop a more systematic approach to early warning.

AI is not one technology

The security implications of AI depend on how and where AI systems are developed and deployed. The Policy Brief distinguishes between four AI modalities:

• Algorithmic AI, which operates through fixed, human-written rules and can shape what information people see.

• Machine-learning AI, which infers patterns from data and can behave unpredictably when conditions or data change.

• Embodied AI, in which computational systems are connected to physical action, such as drones, robots or autonomous vehicles.

• Generative AI, which can produce synthetic text, images, audio and other content and reshape the information environment.

The most significant security risks may not arise from individual modalities alone. They can emerge through interactions between them. The Policy Brief identifies three dynamics in particular:

Layering occurs when weaknesses or errors in one AI system propagate into another modality.

Displacement occurs when controls on one modality shift activity towards another rather than reducing the underlying capability.

Escalation can occur when different AI systems reinforce one another at a speed or scale that makes meaningful human oversight increasingly difficult.

This means that assessing AI systems individually can overlook risks that emerge from their interaction.

Autocratic AI developement is not one threat

A second distinction concerns the geopolitical context. While both China and Russia challenge aspects of the EU’s rights-based approach to AI governance and seek to strengthen state control through AI, they pursue distinct approaches to AI development, ethics and international governance.

Russia places greater emphasis on “traditional values”, resistance to Western norm-setting and military AI, while China combines state-led coordination with a “people-centred” approach focused on socio-economic development and stability. Both approaches can shape international debates about how AI should be governed, but they do so through different strategies and priorities.

Understanding these differences is important for European policymakers seeking to anticipate how AI capabilities and governance models may affect European security.

From risk assessment to early warning

The Policy Brief proposes an AI Threat Index (AITI) and a live dashboard as a practical instrument for monitoring these developments. Rather than treating AI as a single category, the AITI would track AI development across defined high-risk domains and modalities and provide warnings when critical thresholds are crossed.

The Policy Brief also sets out recommendations for actors shaping AI governance and security in Europe and beyond. These include the European Commission and AI Office, national security ministries, standards bodies and AI safety institutions, international organisations, research funders, and foreign policy actors.

The recommendations range from introducing a modality layer into EU AI governance and strengthening technical evaluation and red-teaming, to improving cross-modal risk assessment, supporting research into agentic AI, and monitoring efforts by Russia and China to shape global AI norms.

The central message is straightforward: Europe needs to measure the threat before it arrives.

The goal is not simply to measure how capable AI becomes, but to understand where capabilities create security risks, how those risks may develop, and where early action is needed.