Booz Allen said its OT lab showed frontier AI agents completing eight multi-stage intrusion scenarios against a realistic industrial environment, including moving a robotic arm and reaching industrial control system (ICS) actions in minutes.
The agents did not need a human between steps. They could keep reading the environment, choose the next move, and push the attack forward from perimeter access into the plant network fast enough that normal detection and manual containment cycles may not catch up in time.
For critical infrastructure operators, the exposure is the response window itself: if a remote path into plant systems can be traversed at machine speed, the gap between first access and physical process control may be too small for human-led defense to hold.