Help Net Security reported on 13 October that Booz Allen tested frontier AI models in its Chantilly, Virginia lab. The setup copied a factory: hundreds of IT and operational technology (OT) devices, a SCADA system, and a robotic arm. SCADA is software that lets operators watch and run a plant. The models met all eight attack goals Booz Allen set. They started with only network address ranges. In the last test, one model found the arm, took admin control and moved it. One run took just over 16 minutes from break-in to action in the control zone.
Humans had to approve each step before anything physical happened. Firewalls, segmentation and endpoint tools slowed the models or raised alerts, but the models often tried another route. Help Net Security says they mostly used the devices' own protocols and default passwords. The article does not name the models or say how many runs were made. It does not say how the results would carry over to a real plant. It also notes that none of the recent water and power attacks has been publicly tied to AI. Dragos reported in May that AI helped an attacker map Mexican government networks, with no sign the attacker reached control systems.
For people who run or buy industrial software, the weak points here were old ones: default passwords and one gateway holding live sessions to fourteen devices. What changed was speed, since work that took several specialists days fit into a single run.