The Register Consider this the wake-up call
When it comes to cyberattacks that shut down power grids, drinking-water systems, and other critical infrastructure, AI doesn’t – yet – turn script kiddies into sophisticated, nation-state attackers. But “the potential is there,” according to operational technology (OT) security experts.
"One thing that we've relied on for years is security by obscurity when it comes to utilities," said Steve Swick, former chief information security officer at American Electric Power. "The knowledge needed to do some of that stuff is getting less likely to be necessary," Swick added during a media panel ahead of the publication of Frenos' State of AI in Security report.
AI makes it easier and faster for human attackers to target critical infrastructure, gain access to utilities and manufacturing plants, and perform OT system reconnaissance.
“Basic access can recon documentation, and distill through that documentation very quickly to find out critical design features of your plant and facility,” said Paul Shaver, head of OT at Armadin, during the panel. “I can download the manual for almost any automation controller, parse through that very quickly, and come up with ways that I can exploit a dedicated feature of something that was not secure by design.”
Armadin is Kevin Mandia’s new security startup that uses swarms of autonomous AI agents to simulate real-world cyberattacks. Prior to joining Armadin, Shaver spent six years as the global OT practice leader at Mandia’s earlier company, Mandiant, which Google acquired in 2022.
“I will fly my Mandiant flag for a second and try to steer away from the fear, uncertainty, and death of the world is coming to an end, and every script kiddie is now a very sophisticated nation-state adversary. But I’m not going to say the potential is not there, because the tools are improving such that the access to information is so much quicker.”
At Armadin, and with an AI assist, “I'm very quickly building tools in days that would normally have taken me months,” Shaver said. “If it can do that for me, it's doing something similar for the folks that have a very basic skill set. I wouldn't go so far as to script kiddies yet, but as the models improve, that becomes more of a possibility in this space.”
Plus, as AI lowers the technical barriers to would-be OT hackers, “it advances the timeline,” said Harry Thomas, chief technology officer and co-founder of Frenos, a security company that does AI pentesting for OT environments.
“If a better-than-average script kiddie can now perform attacks,” Thomas said, “You have more threat actors out there. There's an acceleration in timeline, more attacks could be happening, and from a defensive position, you might not know who your threat actor is anymore.”
Testing the power of LLMs to help plan attacks
The Frenos report, published Wednesday, examined the extent to which AI could help someone plan an attack on industrial systems. To do this, it ran several AI models through 16,980 tests between October 2025 and September 2026. These included 13,884 decision attempts, 2,520 planning attempts, and 576 simulated attacks.
The company won’t disclose the models or providers’ names, but said it tested six frontier models, two US open weight models, and two foreign open weight models, on three kinds of tasks: choosing one action, planning the steps to a goal, and working through a simulated scenario.
Each test gave the model background information and a set of attack actions to choose from. Before choosing, each model received details about assets, networks, and known weaknesses, plus the access each action required.
Frenos then checked whether the model’s choice matched the required outcome. Depending on the model, between 6 percent and 47 percent of the choices passed.
These tests found that the strongest three frontier models succeeded in about 90 percent of their planning attempts – meaning they produced a workable attack plan – but these same models weren’t consistent: 60 percent, 62.9 percent, and 74.3 percent of the attack plans passed every repeat.
“A plan can look sensible and still leave out something essential,” according to the report. “The models had information about the task, but some answers skipped a step or assumed access that the plan had not established.”
Meanwhile, the most successful of the four open weight models – a non-US model – achieved a 36.2 percent overall success rate. That figure, however, was still higher than the overall success rates of three other US frontier models, which scored 31.9 percent, 18.8 percent, and 17.9 percent, respectively.
“We're seeing that even though AI models help us accelerate our tooling, accelerate our access and knowledge transfer, they are still prone to hallucinations – aka mistakes,” Thomas said. “We still think human judgment remains essential for these AI models to make the threat actors achieve whatever mission or goal they're trying to achieve.”
In these tests, no humans asked the models to correct their plans, whereas in a real-life attack, a miscreant who understood the target could spot a gap in the AI’s plan, provide more information, and ask the model to revise its answer.
“On top of that, when we were testing the AI models, we used a harness that was pretty simple: Here's some information. Here's the context. Give me an answer,” Thomas said. “A more capable harness could have given these AI models more of an edge.”
Based on the findings, Frenos recommends actions for OT defenders to take, starting with a layered defense. This involves segmenting networks, limiting remote access and user permissions, and testing recovery steps.
“We have our business continuity plans and disaster recovery plans for a reason,” Thomas said. “Let's execute them.” ®
Get AI news in your inbox
Daily digest of what matters in AI.