Booz Allen Tests Found AI Took Control of an OT Robotic Arm in 16 Minutes
Summary
Booz Allen says AI models crossed from an OT network’s perimeter to energize a controller in 16 minutes during lab testing. Separate Claroty research found an AI agent chained known flaws to gain root access on a lab refrigeration controller.
Key points
- Booz Allen says its AI models met all eight test objectives in a mock OT environment, including compromising a SCADA system and energizing a controller.
- The 16-minute test included human approval and unpatched vulnerabilities; Booz Allen estimated it could have taken half as long without human-in-the-loop approval.
- Claroty’s Team82 used an AI coding agent to chain two previously disclosed Danfoss controller vulnerabilities and gain root access on a lab device in 28 minutes.
- Claroty said the vulnerabilities were later involved in an attack on refrigeration systems in France, after the manufacturer had been notified and issued a firmware update.
- The article says the tests did not show an attacker compromising an AI agent already deployed in an operational environment; researchers described that as a scenario organizations should consider.
- Recommended defenses include network segmentation, monitoring for anomalous behavior, and honeypots or canaries to detect or slow attackers.
Article Details
- Event Type
- Laboratory testing of AI models demonstrated rapid compromise of an operational technology environment, including control of a robotic arm.
- Impact
- In Booz Allen's testing, the models completed all eight objectives and took 16 minutes to move from the perimeter firewall through IT and OT networks and energize a controller output. Separate Claroty Team82 research used an AI coding agent to chain two previously disclosed vulnerabilities in a Danfoss refrigeration controller and gain root access to a lab device. The article also reports a recent attack on refrigeration systems in France linked to those vulnerabilities.
People
Jay DeenAssociate principal adversary hunter at Dragos; discussed an AI-assisted intrusion that approached but did not enter a utility's OT control network.Kyle MillerVice president of infrastructure cybersecurity at Booz Allen; discussed the AI testing and its implications for OT security.Sean TuftsField chief technology officer at Claroty; discussed Team82's AI-assisted analysis of refrigeration controllers.
Vendors
Booz AllenConducted by Booz Allen, the government and military contractor, the tests found that in a typical OT environment, commercial AI models could identify misconfigured devices and work their way through multiple layers ofClarotyIn its own research, OT cybersecurity firm Claroty's Team82 used an AI coding agent to analyze Danfoss refrigeration controllers and found three previously disclosed vulnerabilities.Dragosmethods of gaining access, says Jay Deen, associate principal adversary hunter at OT cybersecurity firm Dragos, which reported the incident.