A fifth face? The evolving threat of nuclear terrorism in the age of artificial intelligence

The Nonproliferation Review

Artificial intelligence integration into non-state actor operations threatens to introduce a 'fifth face' of nuclear terrorism, significantly lowering technical barriers to acquiring or fabricating crude nuclear devices. Large language models and generative design tools enable extremist groups to bypass traditional information controls, accelerating the weaponisation of radioactive materials.

Historically, the international nonproliferation regime focused on securing physical fissile material and restricting access to highly specialized scientific expertise. The threat is no longer purely physical. Democratization of advanced machine learning algorithms disrupts this paradigm by automating complex engineering tasks and optimizing chemical synthesis pathways. This digital shift exposes critical vulnerabilities in physical security systems at civilian nuclear facilities, where AI-driven cyberattacks could sabotage reactor cooling mechanisms. Furthermore, autonomous unmanned systems powered by edge AI could be deployed to bypass traditional perimeter defenses, facilitating the theft of radiological sources. Consequently, the International Atomic Energy Agency faces unprecedented challenges in monitoring decentralized digital threats that elude conventional safeguards.

Comment

The convergence of generative artificial intelligence and nuclear security vulnerabilities exposes structural limitations in the Convention on the Physical Protection of Nuclear Material. While the treaty historically focused on securing physical facilities and transport routes, the proliferation of large language models lowers the cognitive barriers to weaponising radiological materials. Extremist groups can now exploit open-source AI tools to synthesize complex enrichment procedures or optimize dispersal mechanisms for radiological dispersal devices. This shift effectively bypasses the traditional intelligence chokepoints that previously relied on tracking the acquisition of specialized scientific literature.

Consequently, this digital evolution places unprecedented strain on the IAEA's Incident and Trafficking Database, which is structurally unequipped to monitor the intangible transfer of AI-optimized weaponization blueprints. Downstream, national intelligence agencies will likely find their traditional surveillance of physical supply chains insufficient as virtual design optimization replaces physical trial-and-error. Ultimately, this regulatory gap shifts the burden of non-proliferation from physical border security to active cyber-interdiction operations conducted by agencies like the US Cyber Command against hostile digital infrastructure.

Strategic Question for Discussion

If generative artificial intelligence tools permanently lower the cognitive barriers to fabricating radiological dispersal devices, how can the Convention on the Physical Protection of Nuclear Material adapt to regulate intangible digital blueprints without infringing on open-source scientific research?

The trajectory indicates that the Convention on the Physical Protection of Nuclear Material will struggle to address intangible transfers, as its legal mechanisms are strictly bound to physical jurisdictions and tangible materials. My assessment is that state parties will increasingly rely on unilateral export controls and offensive cyber operations to disrupt the digital infrastructure hosting these AI models. Consequently, the treaty is likely to remain a baseline for physical security while actual non-proliferation enforcement shifts to dynamic, tech-focused coalitions.

Share your assessment in the comments below.

To read in details click here