
Robots Hacking Robots: AI Can Now Do Most of an OT Attack, Pending a Human Nod
Ourtake
The lab models had to ask permission before moving a robotic arm. Booz Allen's researchers say the models are learning faster than the guardrails.
What happened
Booz Allen Hamilton and Frenos separately tested frontier and open-weight AI models against an air-gapped OT lab network. Booz Allen reported that both frontier models succeeded across all eight attack scenarios, with human approval required before any kinetic action. The researchers said the attacks were often inefficient and noisy, and that some human guidance remained necessary.