FRONTIER AI models operating in OT environments—such as utilities and manufacturing networks—could, in lab testing by Booz Allen, identify misconfigurations and traverse from IT into OT networks to enact physical actions. In one scenario, an AI-driven attacker managed to take control of a robotic arm within 16 minutes, energising a controller in the OT environment after moving through the perimeter firewall and internal networks.
Booz Allen notes this rapid path to a kinetic outcome was the standout finding, even though the attackers did not necessarily cross fully into the OT control network in the observed runs. The tests reportedly achieved all eight objectives, from mapping the environment to exploiting security weaknesses and compromising a SCADA system; researchers also observed that faster execution would be possible if human-in-the-loop approvals were not required.
Claroty’s Team82 separately demonstrated that an AI coding agent could analyse refrigeration-controller firmware and chain two disclosed vulnerabilities to gain root access on a lab device, underscoring how AI can shorten the window to exploit. Experts emphasise that, while there is growing interest in AI-assisted attacks, there is currently limited evidence of widespread OT compromise via autonomous AI agents in actively deployed environments.
Practical responses focus on segmentation and micro-segmentation to slow AI movement, enhanced anomaly-based monitoring to detect unusual AI behaviour, and deception technologies such as honeypots or canaries to deter and reveal intrusions. Operators are urged to pre-emptively map critical paths and harden devices that, if accessed, could yield significant operational impact. 16 October 2026