How AI Is Changing OT Security

AI accelerates both sides of OT cybersecurity. When AI writes the malware, defenders need to understand exactly how capable the threat is and where the real risks lie. Attackers use AI coding assistants to speed up ICS malware development, but functional malware still requires deep engineering knowledge of industrial protocols and physical systems. Defenders gain a more immediate advantage: AI systems that decode raw OT network packets for SOC analysts and automated frameworks that discover and patch zero-day vulnerabilities without human intervention. At S4x26, presentations from Dragos, Panasonic, Darktrace, and Georgia Institute of Technology mapped where the advantage sits today and where cognitive bias risks undermining it.

AI-Generated OT Malware: Faster Development, Same Barriers

When Dragos researcher Jimmy Wylie attempted to recreate FrostyGoop, an ICS malware targeting Modbus devices, using AI coding assistants including Claude, the results showed clear limitations. AI struggles with the nuances of proprietary industrial protocols like DNP3. Generating malware that can disrupt physical processes still requires the ability to build testbeds, understand target systems, and engineer payloads for specific industrial environments.

AI-generated ICS malware accelerates existing attack methods. It does not lower the expertise barrier or create new categories of OT threat. AI can automate vulnerability discovery and assist in reconnaissance, but it does not replace the specialized knowledge required to build and deploy functional OT attacks.

AI Decoding OT Packets for SOC Analysts

SOC analysts monitoring industrial networks face a persistent problem: raw OT protocol packets require specialized knowledge that most IT-trained analysts do not have. Reading Modbus, DNP3, or IEC 104 specifications to interpret intercepted traffic is slow and error-prone.

Rashi Sharma of Panasonic R&D Center Singapore presented PAIEL (Packet Analysis and Insight Extraction using LLM), a system built on a two-stage Retrieval Augmented Generation (RAG) approach. Instead of relying on the LLM’s internal knowledge, which is prone to hallucination, PAIEL pulls context from network protocol standards and maps it to the intercepted packet. The system translates binary data into plain-language analysis: a specific packet is commanding an air conditioning unit to a target temperature of 100 degrees Celsius.

Separately, at the DARPA AI Cyber Challenge, Team Atlanta from Georgia Institute of Technology demonstrated a multi-agent LLM framework that discovered zero-day vulnerabilities in large codebases and automatically generated, tested, and deployed functional patches without human intervention.

Cognitive Bias Undermines AI-Assisted Detection

AI detection tools introduce sociotechnical and cognitive biases that can be more dangerous than the threats they are designed to catch. Nicole Carignan of Darktrace identified three patterns that recur across OT security operations.

Algorithm aversion occurs when analysts distrust AI models after encountering false positives, even when the model outperforms manual detection over time. Automation bias is the opposite: analysts accept AI outputs without verifying them against the underlying OT context. Confirmation bias leads analysts to favor AI insights that match known attack patterns while missing novel techniques. Living off the Land (LotL) attacks, which use legitimate tools and leave no traditional signatures, are particularly likely to be overlooked.

Organizations cannot train away these biases. Detection systems need built-in friction, feedback loops, and accountability mechanisms. Formal human-in-the-loop governance and systematic review of what tooling misses are necessary to prevent over-reliance on automated detection.