AI Models Seize Control of Robotic Arm in 16 Minutes
Booz Allen’s lab testing found that commercial AI models could traverse a mock enterprise OT network from the perimeter firewall through the IT layer into OT and execute a kinetic outcome—energizing an output on a controller—in 16 minutes. The models achieved all eight test objectives, including environment mapping, vulnerability discovery, and SCADA compromise. Kyle Miller, VP of infrastructure cybersecurity at Booz Allen, said the process would have taken roughly half that time without human-in-the-loop approval. The initial runs intentionally left unpatched vulnerabilities in place, so the clock did not include zero-day exploit development.
In a real-world incident reported by Dragos, a small hacking group used two commercial AI models to breach at least nine Mexican agencies in early 2026 but failed to cross into the OT network of a local water and drainage utility. Jay Deen, associate principal adversary hunter at Dragos, stressed that the AI reduced the expertise and time required for large-scale intrusion rather than introducing novel techniques. Separately, Claroty’s Team82 used an AI coding agent to analyze Danfoss refrigeration controllers, chaining two previously disclosed vulnerabilities to gain root access on a lab device in 28 minutes at a cost of about $6 in tokens. Those same vulnerabilities were later exploited in an attack on refrigeration systems in France, two months after Claroty notified the manufacturer.
Sean Tufts, field CTO at Claroty, noted that 70% of companies now use AI in operations and warned that an AI agent with broad permissions could become an additional attack path if compromised. Booz Allen recommends micro-segmentation, anomaly monitoring, and deceptive technologies such as honeypots and canaries to slow AI-assisted attackers and feed them false data. Miller cautioned that automating defenses in safety-critical OT environments will take time to gain operator trust, but detection-triggered perimeter blocking is essential to match AI-speed attacks.
Read the full article
Continue reading at Dark Reading →This is an AI-generated summary. Read the original for the full story.