Industrial facilities, including potential military targets such as nuclear power stations, have been vulnerable to cyber-attacks with real-world consequence ever since their operational technology (OT) systems were connected to the internet via IT systems.
But now artificial intelligence (AI) is moving the attack threat to a new and potentially dangerous level. Booz Allen, a US government and military contractor specializing in intelligence, technology, and cybersecurity, conducted a series of tests proving that the latest sophisticated AI, what the contractor calls "frontier super intelligence", was able to break through the security perimeters of eight scenarios taking place in Booz Allen's isolated OT Cybersecurity Lab.
The frontier super intelligence navigated unfamiliar industrial systems, quickly adapted when blocked and was able to move physical equipment within minutes. For the cybersecurity defenders of industrial facilities, this can mean far less time in which to detect and stop digital access from becoming physical action, which is exactly what happened during the tests.
In a little over 16 minutes, frontier super intelligence models in one controlled test swiftly moved from a perimeter compromise to actions inside an industrial control network. In another scenario, a model located a robotic arm, mapped its protection zones and motion limits, and moved it within minutes. The sophisticated AI models found little difficulty in circumnavigating whatever controls and blocks had been installed to stop them.
"No safeguard provided consistent protection"
"Firewalls, segmentation, and endpoint protection slowed or blocked other actions. The models sometimes switched methods or followed another application path, so no single safeguard provided consistent protection across the evaluation," says Booz Allen.
The tests conducted in the research lab showed that, in the real world, all types of industrial facilities are vulnerable to similar attacks. In one scenario, a model noticed a safety relay repeatedly searching for a missing communications partner. From that single clue, it identified a possible way to impersonate the missing peer. Across all eight scenarios, every defined test objective was achieved.
"The consequences at an operating facility would depend on its architecture, equipment, safeguards, and the access an attacker obtained. Within the lab, the models showed how quickly partial network access could open paths to machinery and industrial processes," says Booz Allen.
While the frontier super intelligence used by Booz Allen in the tests is not yet generally available to ordinary cybercriminals, the speed of development of AI services means that it soon may be. In the case of power stations and other critical facilities, however, the software used by Booz Allen is also likely to be available to potentially hostile nations with the resources to deploy it in the real world of industrial espionage.
