From data centers to power generation and water systems, the AI economy depends on complex industrial environments. Yet many operators still lack the visibility, segmentation and security fundamentals needed to defend them.
OT security insights
- The AI revolution is entirely dependent on physical energy and utility infrastructure, meaning a failure in OT security directly threatens the viability of the AI economy.
- AI-driven tools are rapidly compressing the timeline between vulnerability discovery and exploitation, making it impossible for defenders to rely on traditional, slow-moving remediation cycles.
- The most effective defense against AI-enabled adversaries is the rigorous implementation of security fundamentals like network visibility, segmentation and process-level monitoring.
Every large language model, every GPU cluster running inference at scale, every AI data center humming with activity runs on one thing: electricity. That electricity flows from grids, pipelines, substations, turbines and water systems. Every one of these things is controlled by operational technology (OT). A recent webinar hosted by cybersecurity experts from Dragos, the NSA and Frenos made the case that the AI era is not just a software story. It is an energy and infrastructure story, and the operational technology underpinning it all remains dangerously underprotected.
OT security: The conversation AI is missing
The webinar, titled “Blind to the Blast Radius: Why the AI Era’s Biggest Risk Is the Infrastructure Powering It,” brought together Rob Lee, CEO of Dragos; Rob Joyce, former director of cybersecurity at the NSA; and Harry Thomas, CTO of Frenos. The panel’s central argument was that while the cybersecurity community has focused intensely on securing AI models, data pipelines and enterprise information technology (IT) systems, the operational technology those systems depend on has been left largely unaddressed.
“You can’t have an AI or space or whatever next revolution that you want without energy and infrastructure,” said Rob Lee, who raised the alarm on this issue at the World Economic Forum in Davos earlier this year. “You’ve got to have manufacturing capability, you’ve got to have data centers, you’ve got to have electric power to provide. All of that is the backbone to be able to go and take that opportunity, but also it is the thing that can be disrupted the most. If you want to have your AI revolution, you need operations technology, energy, infrastructure. If you want to keep it very long, you need the security of it.”
Lee noted that the lack of appreciation within the enterprise IT security community for what actually fuels the business is beginning to create dangerous divides and that governments are losing patience.
“A lot of the policy conversations I get into, they’re saying, ‘If the community can’t figure it out, what’s the regulation going to be? Because we don’t have time.’”
AI-accelerated threats: The Mythos moment
A significant portion of the discussion centered on the emergence of advanced AI-assisted vulnerability research tools — particularly a model referred to as “Mythos” — and what they mean for the threat landscape. The panelists were measured but clear that the tools are real, the acceleration is real and defenders need to act now.
“The Mythos curve is real,” said Rob Joyce. “You look at the trajectory we’ve been on for the past year in AI-assisted vulnerability development, agentic red teaming — all of those capabilities are getting better and better. The people that do attacks are really going to be enabled by these technologies.”
Joyce pointed to Microsoft’s most recent Patch Tuesday as a leading indicator. There were 165 new vulnerabilities disclosed in a single month, the second-largest in history, driven in part by LLM-assisted bug discovery. And that was before Mythos.
Harry Thomas offered a historical parallel: “I’d make it akin to when Metasploit came out. People thought the world was going to fall apart. What it enabled was people with lesser knowledge of certain cybersecurity practices to do something they couldn’t do before.”
The consensus among panelists was that AI tools are pulling up the lower end of the attacker bell curve while simultaneously enabling nation-state actors to operate at greater scale and speed.
For OT environments specifically, Lee was careful to separate hype from reality.
“This idea that AI is going to automatically start hacking power grids and bringing down worldwide infrastructure is not represented in anything we see today,” he said. “What I do see is adversaries that are capable being told, ‘Hey, there’s this other stuff you can use your capability against, and here’s a starting point to bring you up to speed on it.'”
He cited a recent incident involving a Mexican water utility, where an attacker using AI tools was guided toward OT systems it would not have otherwise targeted, only to be stopped because the utility had done the security basics.
The boundary is dead … now what?
One of the webinar’s most striking assertions came from Lee regarding the state of perimeter-based security.
“Your boundary is dead for the most part,” he said. “It doesn’t mean don’t do it. You better be putting in a good prevention and segmentation strategy. But it went from your strategy to genuinely just a component of it. If you don’t understand how to detect and respond, if you don’t understand the attack paths that are likely to be taken, you’re going to be left behind very, very quickly.”
Joyce reinforced this with architectural guidance.
“You’ve got to ensure that there’s not a direct path in if one control fails,” he said. “In a world where zero days and end days are more prolific, you’ve got to assume that your firewall, your VPN concentrator, your file transfer box is going to get popped. What are the things that then segment well to keep them from your most important place? And what are all of the logging and overwatch things you’re doing to understand that somebody’s inside your environment?”
Back to OT security basics, with urgency
The panel’s prescription was consistent and unambiguous. Organizations must return to security fundamentals, but with a new sense of urgency. Lee pointed to the SANS ICS Five Critical Controls as a starting framework.
“It’s not that AI has changed those,” he said. “It’s that if you weren’t doing those, you’re in a very bad position now.”
He urged boards of directors to ask a direct question: “Do we have the fundamentals in place against the threats and risks that we actually agree on and care about? If the answer is no, it is extraordinarily disruptive to start with AI in operations. If you have not reached the height requirement to ride the ride and you’re ready to start pushing AI into operations, you could hurt people.”
Joyce introduced the concept of “information arbitrage” as the defender’s core advantage.
“We spent the time to know that network better than the people who owned and operated it,” he said. “You own that network. You get to know what’s normal in that network. An attacker who drops into your network has to figure out where they are and what they want to do. You have this huge advantage.”
He advocated for honey pots and honey tokens as practical, high-signal tools.
“Sprinkle some things around that, when they bump into them, give you strong signal that something is amiss.”
AI in OT security is a tool
The panelists were equally direct about the role of AI in defensive security tooling.
“AI is just another tool,” Lee said. “If you have bad analysts and bad data, AI is not going to do much for you. If you’ve got good analysts and good data, AI is going to be an enabler.”
He cautioned against IT security vendors claiming that AI would suddenly make them relevant in OT environments.
“Your tool doesn’t work today,” he said. “It’s not going to magically work because of AI.”
Thomas predicted that asset owners themselves will increasingly hire AI engineers to harness their own operational data.
“They have decades of data,” he said. “They could take that data with the right senior AI engineer and come up with something very specific to them.”
Lee noted that Dragos had done exactly this, layering a decade of threat research and incident response data onto existing models to dramatically improve their utility for OT-specific use cases.
On the question of autonomous OT security operations centers, the panel was unanimous that we’re not there yet — and perhaps not ever without human oversight.
“I don’t see any capability to be autonomous without human involvement at this point in time,” said Lee.
Joyce agreed: “There’s no way I would let any smart LLM make a decision to implement a firewall rule anywhere and have it understand the implications on the critical processes running. Even if it could, which I doubt, the lawsuits would be endless.”
The societal stakes of OT security
The webinar closed on a broader note about the impact of OT security on society at large. Joyce warned that the pace of AI advancement is outrunning the awareness of most organizations, policymakers and the public.
“Every three months, it crescendos to a new level that I didn’t even expect three months ago,” he said. “The vast majority of people are not having the hands-on keyboard experience to see the phenomenal acceleration this gives the people who use it right.”
Lee offered a blunt summary of where the industry stands.
“Defense is doable and remains so,” he said. “Nothing changes from this discussion. But if you weren’t doing the security fundamentals before, now is the time — today and tomorrow — to go get them right. I see people talking about deception operations and detecting unknown unknowns and what about when they break quantum encryption. Dude, you don’t even have a firewall. You’ve got to start somewhere.”
For the engineers and operators responsible for the physical infrastructure powering the AI economy, the message was that the revolution runs on your systems. Securing them is no longer optional, and the window to act is narrowing fast.
LEARNING OBJECTIVES
- Understand that AI data centers and models are only as resilient as the power and utility infrastructure they run on.
- Learn how AI tools compress discovery-to-exploitation timelines, raising the stakes for vulnerability management and network monitoring.
- Identify why visibility, segmentation and the SANS Critical Controls remain the most effective defenses against advanced AI-enabled adversaries.
CONSIDER THIS
Do you truly possess the “information arbitrage” required to know and monitor your own network more intimately than an adversary using AI-driven tools can map it?
