It was the morning of July 22 in Ashburn, Virginia — the heart of data center alley — when more than three gigawatts of data center load suddenly disconnected from the grid. Voltage disturbances rippled through the region, alarming grid operators. It was the largest data center drop-off event in PJM’s history, representing a sudden loss of about 3% of demand.
The reason was a faulty high-voltage transmission line, which caused a voltage dip and sent the data centers in the surrounding area into protective mode. Phil Tonkin, field chief technology officer at cybersecurity firm Dragos, said it was an example of why data centers can be a big vulnerability for the grid. “If you create something that large on any part of the system, whether it’s producing energy or it stops requiring energy all of a sudden, it can create a destabilizing effect,” he said.
It’s a widely acknowledged concern, even triggering a rare Level 3 alert from North America’s grid reliability watchdog back in May. States are increasingly embracing ride-through requirements. And the industry is rushing to find technical solutions for the problem, such as advanced uninterruptible power supply systems.
But Tonkin is also concerned that the data centers, which have increasingly enormous loads and are getting connected to the grid in very large numbers, could also offer more and bigger opportunities for nefarious actors, operating with more advanced tools than ever before. “What if somebody hacks [the data center] and switches it off?” he said. And that’s just one scenario in which the rapid infrastructure buildout tied to the artificial intelligence boom is exposing the U.S.’ energy system to heightened cybersecurity risks.
Ever since the OpenAI-Hugging Face breach, where a swarm of agents conducted unprompted attacks during some scheduled cybersecurity evaluations of OpenAI models, was revealed this summer, conversation around AI safety has turned to catastrophe.
The breach got public attention in September, when Anthropic researcher Jacob Coxon quit his job, saying people working in AI are seriously concerned that models could go rogue and destroy humanity. A few days later, Anthropic CEO Dario Amodei stoked the flames with an essay encouraging the industry to slow down the pace of AI development to prevent catastrophic damage.
But Leo Simonovich, global head of industrial cyber and digital security at Siemens Energy, is more worried about the vulnerabilities AI has already introduced to the grid than “p-doom” scenarios, a term used to describe AI’s existential threat short for “probability of doom.”
“All that stuff is theoretical, but it feels pretty far away from the reality of the things that we need to tackle today, which can cause a lot of harm,” Simonovich said. “We have an interconnected system…where there are cascading effects that are known. We have aging infrastructure. We have new infrastructure that’s being built.”
“To be very frank, you don’t need the most advanced models to cause damage to the grid,” he added.
Indeed, AI has already changed everything about energy cybersecurity, from what infrastructure is at risk, to how the cyber criminals behave, to how the security itself behaves.
Markus Mueller, field chief information security officer at cybersecurity company Nozomi Networks, said that’s especially challenging in a world with a “much more difficult threat environment” than existed even a year or two ago. He explained that over the past 18 months, events like the war in Iran and rising tensions in the South China Sea have heightened the threat from foreign adversaries. This year, for example, a wave of disruptive cyberattacks against water utilities in the U.S. has been tied to Iranian state-backed hackers.
Meanwhile, changing energy systems are integrating new infrastructure, including more renewables, which “have a completely different risk profile” than large gas or coal power plants. Wind farms, solar plants, and battery sites rely more heavily on remote access, which opens more doors for attackers. Many of these sites are also built or managed by investment firms, which might have a more lax approach to cybersecurity compared to utilities, Mueller added.
A lower bar for attackers
The biggest change in recent years is that AI has made conducting a cyberattack far easier than it used to be. Sean Tufts, field chief technology officer at cybersecurity firm Claroty, said that it used to take a crew of skilled people to conduct cybercrime: from people who knew automation equipment to those who could build malware. A ransomware attack required a money launderer; an attack on a utility required a power plant designer who understood the infrastructure they were looking at. Those were niche skill sets, Tufts said. But “through modern AI, that learning curve has dropped to one chat away.”
Take the December 2015 cyberattack against Ukraine’s grid, which was the first known hack to successfully target utility operations, turning off breakers at three power distribution companies and leaving roughly 230,000 customers without power for several hours in the middle of winter. That attack was conducted entirely by human attackers, who penetrated the network through phishing emails, and then spent several months mapping the system and developing targeted operational attack software.
Ben Sooter, director of agentic AI and micro data centers at the Electric Power Research Institute, explained that those attackers would have likely been faster and more efficient if they had today’s AI tools at their disposal, and not only because AI can write better, stealthier phishing emails.
“Now with AI, reconnaissance is going to happen much faster because…it can run 24/7; it doesn’t have to worry about people needing to go to sleep and eat,” Sooter said. “And then when you find an operational system, reverse-engineering it and creating an exploit is much simpler. The key there is that it definitely lowers the bar for the amount of operational knowledge that you need to make an attack.”
It’s not just power plants themselves that represent a hugely disruptive target. Those who work in the space expect that data centers, too, will become targets. Though they haven’t suffered a publicly known breach yet, Tonkin says the “rapid scale of implementation” and “the sheer complexity of their supply chain” is likely to attract attention from bad actors. Diesel generators, for example, could be used to compromise a data center’s power stability, which in turn could create a destabilizing effect on the grid. “[Diesel generators] tend to have remote connectivity, they’re built as a commodity item, they’ve got built-in vulnerabilities, and they’re very repeatable, so a lot of work needs to go into thinking about how to appropriately secure [them],” he said.
That said, Tonkin has seen most data center companies “thinking about cybersecurity from the very beginning,” all the way at the design stage. Not everyone agrees, though; Mueller at Nozomi noted that “most data centers don’t have a very mature cybersecurity program and lack a lot of controls that we see in other places.”
“Data centers are scary,” Mueller added.
Failing perimeters
Data centers are at risk in part because they are new, and going up fast. But the old stuff that makes up the vast majority of the grid is just as vulnerable, if not more so. As Tufts put it, the entire Earth is criss-crossed with infrastructure and automation equipment that was installed before modern cybersecurity practices.
“We have to replace 100% of our infrastructure,” he said. “Until that’s done, we’re always going to have a gap between best practice and what we can do on equipment my father installed.”
Traditionally, in critical infrastructure assets, operational technology equipment, or OT — the physical machinery and controls that keep electricity flowing, from substation switches to the computers running power plants — has been kept neatly separated from IT for security reasons, so that a cyberattack against a utility’s corporate IT system can’t affect its grid operations.
[AI] definitely lowers the bar for the amount of operational knowledge that you need to make an attack.
Because OT assets are often old and difficult to secure, the industry has tried to protect them by building a secure perimeter around them, such as a firewall.
Those walls put up by energy companies and utilities, however, are becoming easier to compromise. And as a result, third-party cyber firms are just now starting to see the extent of the OT system’s vulnerability, Tufts said. “For a long time, the operators in the field said, ‘Security, keep out.’ Starting about eight years ago, they realized they needed more help,” he added.
The divide between OT and IT created two problems, it has now become clear. First, companies often have no idea whether their systems have been compromised at all. “Having a walled garden where you can’t see anything behind it, threats lived for decades. That’s a real stat,” Tufts said. “We’ve had malware nodes live in client theaters for 15, 20 years at times. We find those on a weekly basis.”
Second, once attackers are past the wall, they have relative freedom of movement.
In the December 2025 cyberattack against Poland’s energy infrastructure, for example, hackers targeted more than 30 renewable energy facilities and two combined heat and power plants. Once they got past the perimeter, “it was game over,” Tonkin said, explaining that the way the sites were interconnected allowed attackers to move from one site to the next.
“We have to further reinforce those perimeters and make many layers — much more like an onion than just one hard perimeter,” he said.
These are problems companies like Claroty, Nozomi, and Dragos are trying to solve for their customers, which include both utilities and other energy companies. And among other tools, they’re actually enlisting AI to help them gain visibility into their systems and segment their networks (meaning creating that proverbial “onion”), among other things. The technology has been a gift to both the criminals and those looking to stand in their way, thanks to its ability to parse through vast amounts of data and single out threats, anomalies, and vulnerabilities quickly.
Simonovich says tailoring AI for OT security specifically is an industry-wide priority at the moment. “Where we need to go is towards an approach where those models have been trained, tuned, and then thoughtfully deployed for the OT environment,” he said. “We’re not there yet. We’re just beginning to think about this problem set.”
A small-player problem
Cyber problems are especially acute for smaller energy projects and companies, which don’t have the funds or the requirements to have solid cybersecurity protocols. There’s some consensus among experts that the largest, most critical utilities in the U.S. have mature and updated security protocols in place, in part because they have to comply with the North American Electric Reliability Corporation’s Critical Infrastructure Protection standards. (NERC is the same watchdog who issued the reliability warning back in May.)
“In North America, we have very strong regulation in the energy sector — I think it’s the best in the world,” Mueller said. “The problem is that the scope has been more towards the traditional larger utilities…If the site isn’t big enough, it’s not going to have requirements from a compliance standpoint.”
Attacks on smaller sites are less problematic on their face, but if multiple sites are targeted all at the same time, which is what happened in Poland, they can still be dangerous for grid stability, Mueller added — and they’re disruptive nonetheless.
Outside the energy realm, in 2026 hackers linked to Iran coordinated attacks against water utilities across at least 12 U.S. states and at least 100 facilities. While none of the attacks caused serious damage, they signaled an obvious vulnerability within the country’s critical infrastructure system, as well as the extent of Iran’s willingness to exploit it.
“They’re probing our systems every single day,” Tufts said, explaining that, while other foreign actors want to destabilize (namely Russia) or quietly gain access (China), Iran mostly wants to impose costs on the U.S. “They are present on all of our electric utilities’ firewalls, and currently we’ve seen five waves of Iran’s attacks… always focused on OT device-level vulnerabilities that can be taken advantage of from a remote location.”
Water utilities tend to have limited budgets and personnel, and internet-facing devices that have not necessarily been protected as well as they could have, he added, which makes them easy targets.
The attacks have kickstarted a wave of awareness of cybersecurity threats even in smaller players, says Tonkin, whose company Dragos gives its software for free to water or electric organizations with less than $100 million in annual revenue as part of its Community Defense Program. Ever since the attacks, Dragos has seen a “huge increase” in interest in the program, which was launched in 2023 to provide the kind of OT security many small companies can’t afford.
The program is one of several that has emerged to foster cooperation in countering the increase in cyberrisk. Others include OT-CERT, a free membership program for small and medium industrial organizations, which is also run by Dragos and which Tonkin describes as a “community of knowledge transfer,” and Project Watershed 250, a White House pilot launched this summer in Texas, which gives water and wastewater utilities free cybersecurity and AI tools.
“A company that maybe has five people looking after the engineering for a local water company… is not going to have a dedicated cybersecurity person,” Tonkin said. “They’re lucky if they even have a dedicated IT person.” So in a moment when the risks are coming faster than ever, and when the grid is more complicated than ever, these collaborations are ideally a way to get holes patched faster than an individual company could on their own.
“While everybody is talking about future scenarios, there’s a set of problems that are known today, and a set of capabilities that can help us get better,” Simonovich said. “We need to deploy faster — and at scale.”
