AI Models Successfully Breach Industrial Control Systems in Booz Allen Tests

• •AI models breached all defined attack objectives in Booz Allen's tests.
• •The tests demonstrated AI's ability to exploit vulnerabilities in industrial control systems autonomously.
• •Defenders may struggle to respond effectively due to the rapid pace of AI-driven attacks.

Booz Allen Hamilton's operational technology lab tested advanced AI models against an isolated industrial control testbed, achieving all eight defined attack objectives. The models autonomously mapped the network, identified vulnerabilities, and executed commands on various operational technology devices, including robotic arms and SCADA systems. In one instance, an AI agent took control of a robotic arm in just over 16 minutes, demonstrating a significant speed advantage over human defenders. The tests revealed that the AI could adapt to failures, such as targeting the wrong SCADA interface, and still execute successful exploits. The models operated without any source code or engineering documents, relying solely on the provided network boundaries. The findings indicate that organizations lacking foundational OT cybersecurity practices may be vulnerable to such rapid and sophisticated attacks.