Report warns AI agents can already execute devastating attacks on critical infrastructure
A Booz Allen Hamilton study tested eight scenarios in an operational technology lab and found that advanced AI models could achieve all objectives, including moving a robotic arm and breaching industrial control networks. The firm says defenders may have only minutes to detect and block such autonomous attacks. It calls for more industry testing and stronger cybersecurity defenses across critical infrastructure.
Booz Allen Hamilton created an OT testing environment to gauge autonomous AI attack capabilities. It ran eight scenarios in a multi-vendor setup modeled on a manufacturing site, split into enterprise, industrial DMZ, plant operations, and production zones. Equipment included PLCs, HMIs, SCADA, engineering workstations, a variable-frequency drive, sensors, and a robotic arm.
The two advanced models, whose identities were withheld, got no source code, engineering documents, or specialized OT/IT guidance. They still achieved every scenario goal. One moved a robotic arm within minutes; another advanced from perimeter access to industrial control network actions in just over 16 minutes. Human approval was required before exploitation or potentially damaging actions.
If such autonomous capabilities spread, operators of water, power, manufacturing, and other essential services could face compressed response windows, potentially disrupting daily necessities. Security teams and engineers may need faster detection, stronger segmentation, and human oversight. The public could be affected indirectly through outages or unsafe physical processes, though the study was lab-based and does not prove real-world attacks are imminent.