The average US reactor is older than the web, and patches come quarterly

The Verge asked cybersecurity specialists whether rogue AI could take down the power grid, and got a different answer than the question expected: they are still more worried about people holding these tools than about agents acting alone. The reason is not philosophical. It is a gap between two clocks.

Start with the hardware. The average nuclear reactor in the United States is about 44 years old, which puts its construction around 1982, seven years before the web existed and before most of these plants had any reason to expect a network connection. They were eventually connected anyway. Some of the firms that built the equipment still running in the power sector no longer exist, so certain devices have nobody left to write a patch for them at all.

Two clocks that do not match

The three numbers that frame it:

• The average US nuclear reactor is about 44 years old.
• Operational technology systems may accept updates once a quarter or once a year.
• OpenAI pledged $1bn on 3 September toward defending critical infrastructure.

Now put the second number against an attacker. An operational technology system on a quarterly cycle that learns of a flaw the day after its window closes runs exposed for up to 91 days; on an annual cycle, up to 365. Sophie McDowall of the Foundation for Defense of Democracies told The Verge the real change from AI is that adversaries move faster while defenders struggle to match the pace. That is the same statement as the patch calendar, said in words.

What the tooling actually adds is knowledge, not autonomy. Joshua Corman of the Institute for Security and Technology put it plainly: an attacker may not know OT protocols or OT networks, but the LLM has read the manuals and does know what to do. That moves the threat list down from disciplined nation-states to anyone with intent. Rob Denaburg of the American Public Power Association, which covers utilities in 2,000 municipalities, adds the practical half: chain or no chain, stopping the attacker at one point still stops the attack.

“It's literally any sociopath that wants to attack is now more powerful than they used to be. This has been a force multiplier and continues to grow.”

— Joshua Corman, Institute for Security and Technology, quoted by The Verge

Joshua Corman, Institute for Security and Technology, to The Verge

A billion against a trillion and a half

OpenAI pledged $1bn on 3 September to subsidise models that help defend critical infrastructure, saying AI-enabled attacks will become far more widespread in the coming months. Set that against what the same industry is spending to build: the hyperscalers issued $132bn of debt this year and owe $1.5tn on compute contracts. The defensive pledge is 0.76% of one and 0.067% of the other. McDowall's line about offering support for a problem you are partially causing has a ratio attached to it.

Her other point is the regulatory one. Nuclear technology and hazardous materials come with guardrails on research itself; AI does not. Yesterday Nvidia's chief executive put the risk at zero and said no new rules are needed. The utilities are not waiting for that argument to settle. Corman says operators are reaching a simpler conclusion: if we cannot protect it, disconnect it. For infrastructure built before the network, that may be less a retreat than a return.

Informational material, not investment advice. Expert comments and infrastructure figures are as published by The Verge on 20 September 2026; the comparisons against AI capital spending are ours.