Consulting firm Booz Allen Hamilton conducted tests where AI agents successfully compromised a simulated industrial environment. The agents performed eight separate attack scenarios, achieving their objectives in every instance without prior access to source code or engineering documentation.
The laboratory environment replicated a typical production facility, featuring programmable logic controllers, SCADA platforms, human-machine interfaces, and robotic hardware. The setup incorporated diverse vendor equipment and software to mimic the technical complexity of real-world industrial systems.
AI agents autonomously identified vulnerabilities and planned their actions, though they were restricted by requirements to obtain human approval before executing potentially harmful commands. The agents demonstrated the ability to adapt their strategies, such as modifying attack plans when initial attempts to breach SCADA interfaces failed.
In one specific test, an agent gained administrative access to a robotic arm and manipulated its movement after mapping safety zones. Another scenario saw an agent breach the perimeter and infiltrate the internal industrial control network within 16 minutes by exploiting network configuration errors.
