Artificial Intelligence Is Not Yet Ready for Industrial Control System Cybersecurity

When Google Duplex first spoke on a restaurant phone line, it sounded almost human. The demo showcased what AI can do today: execute narrow, scripted conversations with uncanny fluency. Yet the same technology, now available in 43 U.S. states for Google Pixel users, highlights the limits of today’s AI—capable of routine interactions but still far from tackling the complexities of industrial control systems (ICS).

Other high‑profile AI projects underscore this divide. IBM’s Project Debater can construct persuasive arguments, but its voice remains robotic and its understanding is abstract. Both examples illustrate that the most advanced AI comes from companies with deep pockets and massive data sets—yet even they caution that algorithms may fail, data can be biased, and improper practices can erode trust.

Marketing narratives often paint AI as a cure‑all for business challenges, promising to make sense of vast amounts of industrial data and to fortify the security of control systems. Industrial analytics, when applied to machine data, drives the convergence of OT and IT and fuels the Fourth Industrial Revolution, according to the Industrial Internet of Things Analytics Framework from the Industrial Internet Consortium.

Cybersecurity veteran Jason Haward‑Grau, CISO of PAS Global, notes that “robotic process automation is probably far more interesting, from an AI perspective, than AI is in security.” He argues that vendors can claim they offer AI for every problem, but businesses should first ask, “What does my organization need?” and then determine whether AI is the answer.

Threat levels remain high. A 2018 Kaspersky study found that 49 % of 321 industrial respondents reported at least one attack per year—an under‑reporting likely underestimating the true figure.

Defining AI is itself a philosophical challenge. Technology writer Jaron Lanier warned in 2016 that without a measurable baseline, “you are off in fantasy land.” In practice, this means many systems labeled “smart” lack empirical grounding.

One concrete use case is malware or anomaly detection. With a reliable network baseline and robust machine‑learning algorithms trained on sufficient data, AI can flag threats quickly and reduce false positives. This is especially valuable given the cybersecurity talent shortage.

However, success depends on data relevance. If the system is unaware of certain business processes, it may generate false alarms. Moreover, training on flawed data can cause malware to be misclassified as benign. Adversaries may even reverse‑engineer model features to evade detection.

Industrial environments present additional hurdles. OT devices often do not use standard IT protocols, and legacy control buses (e.g., a 25‑year‑old system) are difficult for AI to interpret. Haward‑Grau illustrated this with a refinery example: while the 500 IT devices can be monitored with network tracking, the 28,500 OT endpoints from multiple vendors (ABB, Schneider Electric, Siemens, Yokogawa, Philips, GE, Honeywell) present a fragmented data landscape that complicates AI translation and defining “good” behavior.

The shift from “it’s only a matter of time before we’re breached” to “we’re already compromised” amplifies the complexity of establishing normal network behavior. An IBM‑backed 2018 study showed it takes an average of 197 days for enterprises to detect a breach—time that can erode the reliability of machine‑learning models trained on evolving network topologies.

AI holds promise for industrial cybersecurity, but adoption should start with well‑defined use cases that limit data complexity. As E. F. Schumacher said, “It takes a touch of genius—and a lot of courage—to move in the opposite direction.” For industrial leaders, that means beginning with small, manageable projects before scaling AI across their entire OT landscape.