Frontier AI models completed all eight industrial-control attack tests

Booz Allen tested frontier AI models against factory systems in an isolated lab in Chantilly, Virginia.

The testbed included hundreds of information technology and industrial devices, real controllers, and a robotic arm.

The models completed all eight attack objectives, from mapping the network to gaining control of production systems.

In one run, a model reached the control zone in just over 16 minutes, while operators had to approve any physical action.

The test showed how quickly AI can combine familiar attack methods, including using default passwords and writing exploits, rather than inventing new ones.

That speed matters because it can sharply reduce the time and expertise needed to find weaknesses in systems that operate factories and other critical services.

Booz Allen said the models’ actions were confined to its lab testbed, while separate real-world attacks on utilities have not been publicly attributed to AI.