Specialists Say AI Mainly Speeds Up Human Attackers on Power Grids

TL;DR

• Cybersecurity specialists interviewed by The Verge say AI tools are mostly making existing human hackers faster and more capable, not creating an entirely new kind of threat to power grids.
• At the same time, OpenAI says its newest model, GPT-6 Astra, crossed the highest risk category in the company’s own safety rules after it found two previously unknown security flaws on its own during testing.
• OpenAI is also promising $1 billion in discounted cybersecurity help for utilities and other infrastructure operators, but it hasn’t said what exactly the discount covers or how it compares to the normal price.
• A separate research group found that no cyberattack on a power grid, AI-assisted or not, has ever caused catastrophic damage, because coordinating an attack across many targets is harder than the hacking itself.
• An industry breach report found human mistakes, like stolen logins and unauthorized use of AI tools at work, remain a leading way hackers get into systems generally, though the data does not break this down for utilities specifically.

What happened

Three cybersecurity specialists, interviewed separately for a story in The Verge, republished in full by Bytes Europe, described AI’s effect on power-grid security in similar terms: it is not introducing a new kind of attacker so much as making the attackers who already exist faster and more capable.

Joshua Corman, an executive at the Institute for Security and Technology, told The Verge that AI tools raise the capability of attackers at every skill level, including people acting alone with little technical background. “Any sociopath that wants to [attack] is now more powerful than they used to be,” he said. “A bad-actor human can use these tools to be better than they naturally would be.” He also described the grid’s underlying exposure as long-standing rather than new: “We were always prey. We were just kind of surviving at the appetite of our predators.”

Sophie McDowall, a researcher at the Foundation for Defense of Democracies’ Center on Cyber and Technology Innovation, made a similar point: “The true difference from AI is that it’s letting adversaries move more quickly.” She also argued that AI companies selling cyber-defense products are, in effect, marketing a fix for a problem their own technology helps create: “They’re offering support for a problem that they are partially causing.”

Rob Denaburg, who runs cybersecurity programs at the American Public Power Association — a group representing utilities across roughly 2,000 municipalities — told The Verge that utilities should not treat AI-driven intrusions as needing a wholly separate defense from ordinary cyberattacks: “AI or not, it is at the end of the day, still a cyberattack.”

That framing sits alongside a different kind of evidence: OpenAI’s own internal risk test. According to SecurityWeek, OpenAI’s Preparedness Framework — the company’s internal system for deciding how much to restrict a model before release — has a “Critical” tier for cybersecurity, defined as a model able to independently find and exploit unknown flaws across many hardened real-world systems, or carry out an entire novel attack from just a high-level goal. OpenAI’s newest model, GPT-6 Astra, became the first of its models to cross that tier, after it autonomously discovered two previously unknown security flaws during evaluation and scored a perfect result on ExploitBench, a benchmark that measures whether a model can turn a known flaw into a working exploit. OpenAI said it would still limit rollout of Astra’s full cybersecurity abilities at launch, giving early access to a selected group of testers before wider release through a program called Daybreak Blue: “Full cybersecurity capabilities will not be widely available at launch,” the company said.

OpenAI paired that disclosure with a separate pledge, reported by SecurityWeek: a program called Daybreak for Frontline Defenders, offering $1 billion in subsidized access to its most advanced cyber-focused models, plus training and technical help, for under-resourced defenders including water utilities, electric utilities and local governments, with US organizations prioritized first. SecurityWeek’s own reporting noted that OpenAI did not disclose what exactly the subsidy covers, or how the subsidized cost compares to the price a utility would otherwise pay.

Separately, a Centre for the Governance of AI (GovAI) research paper published July 30, 2026 found that no cyberattack on a power grid to date — AI-enabled or otherwise — has approached catastrophic scale. The paper concluded that the main obstacle to a hypothetical large-scale AI-enabled grid attack is the practical difficulty of coordinating action across many separate targets, describing the binding constraint as “the operational capacity to coordinate diverse capabilities in attacks on dozens to hundreds of targets,” rather than the AI’s technical skill. The researchers characterized a genuinely catastrophic attack as involving “four orders of magnitude more disruption than any grid cyberattack recorded to date.”

On the human side of the equation, Verizon’s 2026 Data Breach Investigations Report — its 19th edition, covering roughly 31,000 security events across breaches from November 2024 to October 2025 — found that exploiting software vulnerabilities overtook stolen login credentials as the single most common way attackers got in, accounting for 31% of breaches. The same report found that employees’ use of unapproved “shadow AI” tools at work tripled in a single year, from 15% to 45% of employees. Within the utilities sector specifically, DeepStrike, an industry cybersecurity research site, reports that Verizon’s data found social engineering, system intrusion and basic web-application attacks together accounted for 94% of recorded breaches.

What this means (and what it does not)

Taken together, the interviewed specialists’ framing and the GovAI research point the same direction: the more immediate risk to power grids described in this reporting is human attackers becoming faster and more capable with AI tools, not AI itself acting as an independent attacker. The American Public Power Association’s preference for treating AI-driven incidents like ordinary cyberattacks also serves its members’ interest in relying on existing defenses rather than costly AI-specific rules — a reminder that even measured framing can serve an institutional interest.

None of this means AI poses no elevated risk. OpenAI’s own framework — again, the company’s internal classification, not an independent grid-specific test — puts its newest model in the top cybersecurity risk tier precisely because it can find and exploit unknown flaws on its own. OpenAI benefits from that disclosure on two sides: the classification supports its safety credibility, while its billion-dollar defensive pledge markets its own AI products to the same infrastructure operators it is warning about. And the GovAI finding that no attack has yet been catastrophic is a statement about history to date, not a guarantee about what happens next.

What we still do not know

The Verge’s original article could not be independently retrieved directly; its quotes and details were instead verified against two full-text syndicated republications that matched each other verbatim, rather than against the outlet’s own site.

No source found breaks down which specific human errors — misconfiguration, credential reuse, delayed patching, phishing susceptibility — occur most often in documented energy-sector incidents specifically; the breach-vector figures cited above describe utilities in aggregate, not a taxonomy of error types. Similarly, no source found gives a direct, energy-sector-specific numerical comparison between human-caused breaches and AI-driven attacks. Whether GPT-6 Astra’s demonstrated ability to find and exploit flaws has ever been tested against energy-sector control systems specifically is not addressed by any source found; OpenAI’s disclosures describe benchmark and lab conditions, not grid targets. No source found evaluates, with data, which training or technical measures have actually reduced human-related vulnerabilities at utilities. Whether any AI-driven cyberattack has already caused real-world damage to energy infrastructure is not established either way. McDowall’s statement that AI companies are “partially causing” the problem they sell fixes for was not tied to any specific company, tool or incident in the reporting reviewed. And the real value of OpenAI’s $1 billion Daybreak pledge to any given utility remains unclear, since OpenAI has not disclosed what the subsidy covers or how it compares to full price.