AI Agents Moved Robotic Arm In Booz Allen OT Tests

Booz Allen tested frontier AI models in an isolated operational technology (OT) laboratory and found that the models completed every defined objective across eight scenarios. The findings were presented last week at an industrial control systems cybersecurity conference in Nashville, before the report’s public release.

The models received no source code, engineering documents or specialized guidance about the systems. The laboratory was designed to represent OT environments, although it may not match every enterprise installation.

The tests combined tasks that usually involve several people and stages, including asset discovery, device research and attack planning. In one scenario, the models moved from compromising a perimeter system to taking action inside an industrial control network in slightly more than 16 minutes.

In another scenario, a model identified a robotic arm, mapped its protection zones and movement limits, and moved the equipment within minutes.

Existing security controls blocked or delayed some actions. No single safeguard consistently stopped the models from completing their objectives, according to the report.

“We have long understood that non-kinetic cyber actions can produce kinetic effects,” Booz Allen researchers wrote. “What has changed is the speed and precision with which these models can help produce those kinetic effects.” The report added that the systems reduce the protection provided by “security through obscurity,” in which organizations rely on unfamiliar or undocumented OT environments being harder to attack.

Booz Allen recommended that OT owners make systems harder to traverse after an initial breach, remove implicit trust between systems, detect consequential industrial changes quickly and limit what an attacker can affect through cyber or physical access.

The report calls the technology “frontier super intelligence.” That term has a separate meaning in AI research. Oxford philosopher Nick Bostrom defined superintelligence in 2014 as an intellect that greatly exceeds human cognitive performance across virtually all areas of interest. The report appears to use “super intelligence” to describe advanced current AI systems rather than a level beyond artificial general intelligence.

The terminology follows a Sept. 29, 2026, executive order from President Donald Trump directing the executive branch to replace “Artificial Intelligence,” or AI, with “Super Intelligence,” or SI. Booz Allen had not clarified its use of the term at the time of publication.