AI Models Breach a Simulated Factory Network in 16 Minutes

The Operational Technology Lab at Booz Allen Hamilton conducted a test of two unnamed frontier AI models in a multi-vendor simulated manufacturing environment, according to a report by The Register. The company stated that both models achieved the target in all eight test cases.

The actions performed by the models included mapping Supervisory Control and Data Acquisition (SCADA) systems, altering values in industrial processes, spoofing a safety sensor device, and moving a collaborative robot arm on the production line, according to the test findings.

In one instance, the model moved from perimeter breach to executing actions inside the industrial control network in just over 16 minutes, which the report describes as faster than human incident response teams can keep up with. This type of attack targets critical facilities such as factories and power plants.

This development particularly concerns industrial operators, as air-gapping operational networks from internet networks was previously considered a key safeguard, but the test indicates that this safeguard alone is no longer sufficient. These findings are pushing critical infrastructure companies to reconsider their defense plans and monitor unusual behavior within operational networks.

What Do These Terms Mean?

Operational Technology: Systems that manage machinery and physical processes in factories and power plants, differing from office information technology systems.

SCADA Systems (Supervisory Control and Data Acquisition): Software that supervises industrial equipment, collects its data, and issues commands to it.

Frontier Model: One of the most powerful available AI models, used here to test its level of risk to critical systems.