How nuclear security can help bring the US and China together on AI

Bulletin of the Atomic Scientists | Stephen Herzog

The United States and China could use artificial intelligence to protect civilian nuclear materials and facilities from theft or sabotage, according to Stephen Herzog. He proposes a six-month pilot agreement using simulated data and digital twins, building on the 2016 Beijing nuclear security training centre developed with Sandia National Laboratories.

AI could help attackers reconnaissance nuclear sites or impersonate personnel, but it could also combine video, badge records, maintenance data and security exercises to flag anomalies. Herzog links the proposal to commercial stakes: Google backs Kairos Power reactors and Amazon supports X-energy in Washington state. Following their September summit, Trump and Xi agreed to AI dialogue, with the next exchange planned for November. Nuclear security offers a limited area where cooperation may begin without military AI restrictions. The stakes are shared. A radiological release would cross borders. Chornobyl showed that in 1986.

Comment

The proposal turns on command and coordination rather than technology alone. Nuclear facilities already generate separate streams of security information: perimeter alarms, badge records, maintenance histories and red-team findings. Herzog’s argument is that AI could fuse these streams, while human specialists retain authority to judge whether an anomaly constitutes a threat.

The 2016 Beijing nuclear security centre, built with Sandia National Laboratories, shows that such coordination has existed before. That cooperation did not require either side to accept limits on military AI; it rested on a shared interest in protecting nuclear materials. A new pilot could similarly separate civilian physical protection from the harder questions of AI-assisted proliferation and nuclear command-and-control.

The precedent of Anthropic’s work with the US National Nuclear Security Administration suggests a workable division of labour: developers supply analytical tools, operators supply operational context, and government bodies retain deployment and enforcement authority. The obstacle is political. Neither Washington nor Beijing endorsed the military-AI declaration at the 2026 Responsible AI in the Military Domain Summit in A Coruña, and both may fear that shared methods strengthen a rival.

Strategic Question for Discussion

If a US-China nuclear-security pilot using simulated facilities produced useful results, would shared operational procedures or independent national deployment authority matter more for sustaining cooperation?

The pattern suggests that independent national authority would matter more, because each government must retain control over sensitive facilities and enforcement. Shared procedures can build technical confidence, as the 2016 Sandia-Beijing centre showed, but they cannot remove sovereign control over nuclear security. My assessment is that a narrow, simulated pilot could endure precisely because it leaves deployment decisions and military AI policy outside its scope.

Share your assessment in the comments below.

To read in details click here