AI risk watchHappening now
Frontier AI agents with no knowledge of the plant met every attack objective in Booz Allen's OT lab and moved a robotic arm
In Booz Allen's isolated industrial lab, frontier models given no source code, documents or OT guidance ran eight attack scenarios against PLCs, HMIs, SCADA and a robotic arm and met every objective. One run reached actions inside the control network just over 16 minutes after a perimeter compromise; another found the robotic arm's interface, mapped its motion limits and moved it. The agents chained discovery, device research and attack planning, adapted when blocked, and no single safeguard consistently stopped them. Booz Allen presented the work at SecurityWeek's ICS Cybersecurity Conference in Nashville.
Why it matters
Not knowing how an industrial network is built has protected a lot of them; an agent that reads the device documentation as it goes removes that protection in minutes.
What you can do
- Segment OT from IT and from the internet so that a perimeter compromise is not a path to the controllers.
- Remove implicit trust between devices, and alert on configuration and setpoint changes within minutes.
- Limit the physical effect any one compromised system can have, with interlocks that software cannot override.