Booz Allen: SI models completed all eight attack scenarios on a simulated factory, including control of a robotic arm
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
In the lab, SI agents can reach industrial equipment, so factory defences have to be ready in advance.
According to Booz Allen Hamilton, in an isolated lab simulating a plant with equipment from different vendors (programmable controllers, operator panels, a SCADA platform, sensors, a collaborative robotic arm), frontier SI models, whose names were not disclosed, achieved the objectives in all eight scenarios; one path took just over 16 minutes. The models were given no source code, documentation or hints about industrial systems. Humans approved exploits and any action with a physical effect, and the results do not mean attacks on operating facilities. In one test the model found the arm, identified its protection zones and motion limits and moved it within minutes; in the SCADA test the models corrected a mistake about the operator-interface version and found active connections to 14 devices. Some safeguards worked: a controller rejected a stop command, and a motor in local mode could not be started remotely. Company vice president Kyle Miller, as relayed by 3DNews, said that SI agents act with such speed and persistence that they can target organisations without industrial cybersecurity methods.