A Booz Allen report, presented at SecurityWeek’s ICS Cybersecurity Conference in Nashville, details how frontier AI agents performed in an isolated industrial‑control (OT) lab. The study used AI models with no prior knowledge of the target OT environment—no source code, engineering documents, or advanced OT guidance—and tested eight scenarios.
In these tests, the agents achieved every objective, including asset discovery, device research and attack planning, and, notably, moving from a perimeter compromise to actions inside an industrial control network in just over 16 minutes. In one test, a model located a robotic arm, mapped its protection zones and motion limits, and moved it within minutes. The work underscores how autonomous AI can operate across both cyber and physical domains at speed unattainable by human intervention.
The report cautions that existing defenses slowed some actions but did not consistently halt the autonomous models from achieving their objectives. While the terminology around “super intelligence” is debated, Booz Allen uses it to describe increasingly capable AI rather than a clearly defined new form of AI. The authors emphasise that autonomous agents can produce kinetic effects via non‑kinetic cyber actions and that OT environments cannot rely on security through obscurity.
They urge operators to focus on making environments harder to traverse, removing implicit trust, detecting consequential industrial changes quickly, and limiting what attackers can affect after access, whether cyber or physical. The overarching message is that autonomous AI is converging cyber and physical security, and proactive preparation is essential as autonomy grows.