Give an AI Industrial Protocol Tools and Watch What It Destroys

The speakers start from a familiar problem: after an industrial attack makes the news, defenders can read the incident report but cannot safely replay it. Scanning real OT devices can crash them, and most ICS malware is targeted, so samples are rarely public. What teams can usually test is a command-and-control address, which tells them little about whether their detections would catch the actual behavior.

To address this, they present mrhOTshOT, an open-source adversarial emulation framework. It covers the full kill chain rather than only the OT payload: an exploitable router as initial access, lateral movement into Windows and engineering systems, and then protocol-native process manipulation. Ten simulated industrial processes run across protocols including IEC 104, Modbus, S7comm, OPC UA and TriStation, each with physics that matches its real process. The demos reproduce Cosmic Energy-style and Stuxnet-style chains, including a case where the operator display shows normal values while the physical process is driven toward failure. Each attack can be exposed through Caldera or an LLM agent that turns a vendor report into a new scenario.

The practical takeaway is to use the framework to validate firewall rules and detection engines, with Suricata shown as the example, on a monitored lab network. The authors stress that the protocol libraries are deliberately insecure, not meant for production, and that the agentic feature is still in beta.