OT / ICS · 8h ago
Booz Allen said frontier AI models cleared 8 of 8 OT attack goals in an isolated manufacturing testbed in Chantilly, Virginia, including reaching a robotic arm, taking administrative control, and moving it in minutes. In one run, the model went from breaching the perimeter to operating inside the control zone in just over 16 minutes.
The models were given only subnet boundaries, then worked out the plant layout, matched devices to weaknesses, wrote exploits, pivoted through the control stack, and in one case used authenticated SCADA sessions to make operator screens lie while sending real commands. The testbed used real controllers, a SCADA platform, sensors, a motor drive, and layered zones meant to mimic a factory.
For defenders, the point is the clock: segmentation, monitoring, and human-in-the-loop response buy less time if recon, exploit development, and lateral movement happen automatically. In plants where a single SCADA foothold can reach multiple controllers, the exposure is not a lab trick but the speed at which control-zone compromise can now arrive.
2 sources covering this story
AI Agents Navigate Unfamiliar OT Systems and Move Equipment in Minutes, Booz Allen Finds
In Booz Allen’s isolated OT lab, AI models met every objective across eight scenarios and reached an industrial control network from a perimeter compromise in just over 16 minutes.
Frontier AI cleared every OT attack goal Booz Allen set in its lab - Help Net Security
Booz Allen’s lab test shows the AI OT security threat is real: frontier models cleared 8 of 8 attack goals and moved a robot arm in minutes.
Part of the PlainSec briefing for 2026-10-14