What If?

Asking the question…

What if the attack on the food supply was not launched by a human?

What if, instead, it came from autonomous AI agents that decided refrigeration was a poor use of electricity and quietly diverted that power back to the data centers that keep them running?

We are not saying that is what happened. We’re just posing the question. What if?

So far the evidence points toward some kind of cyber attack. However, the events of the last three months should make us start asking questions. It gives us a very real preview of how the scenario would unfold, and the preview should bother all of us.

What Actually Happened

In the last week of August, at least six U.S. military commissaries lost refrigeration almost simultaneously: Fort Huachuca in Arizona, F.E. Warren Air Force Base in Wyoming, Naval Station Newport in Rhode Island, Holloman Air Force Base in New Mexico, Little Rock Air Force Base in Arkansas, and Fort Irwin in California.

Meat, dairy, produce, and frozen goods were pulled from shelves. The Pentagon is now treating the failures as a suspected cyberattack.

On October 12, hackers took control of refrigeration systems at hundreds of French supermarkets. They remotely triggered defrost cycles, pushed units to roughly 122°F (50°C), and in some stores started fires.

Then on the night of November 1, it spread. Supermarkets in Sweden, Belgium, Luxembourg, Ireland, Estonia, the Netherlands, Finland, and Germany reported the same pattern. A store in southern Sweden found its freezers set to 122°F (50°C). Roughly forty stores were hit across Belgium and Luxembourg. Thirteen stores in Ireland pulled chilled and frozen food as a precaution. Retailers filled shopping carts with spoiled product and threw it away.

There was no ransom demand, no confirmed attribution, and no claim of responsibility that investigators have been able to verify.

The Boring Explanation Is the Scary One

Investigators and researchers are pointing at refrigeration controllers from vendors that were connected to the public internet, often still running default passwords, often outside the retailer’s central network.

That is the mundane truth. Nobody needed a superintelligence to do this. They needed a search engine, a default login, and a willingness to turn a dial.

And that is exactly why the what-if matters.

What If the Next One Isn’t Run by a Person?

We are not suggesting an AI caused these attacks. The evidence points to human attackers exploiting neglected equipment. But the details deserve a second look, because they describe precisely the kind of work AI agents are becoming very good at.

Consider the shape of the campaign: thousands of exposed devices found and catalogued, logins attempted at scale, commands pushed to hundreds of sites in a single night, defrost cycles stretched just long enough to slip past alarms, and no ransom note. That is a reconnaissance and execution workload. It used to take a team. It increasingly does not.

In November 2025, Anthropic disclosed that a state-sponsored group had used its Claude Code tool to run an espionage campaign in which AI performed an estimated 80 to 90 percent of the tactical work, with humans stepping in at only a few decision points. Earlier this year, Dragos documented an AI-assisted campaign against Mexican government agencies that went on to probe a water utility’s control systems. And in October, Booz Allen reported that frontier models reached all eight attack objectives in an isolated industrial control lab, including manipulating physical controllers, while humans were required to approve each physical action.

Put those together and the realistic near-term risk is not a rogue AI that decides on its own to spoil the food supply. It is a small, unsophisticated group, or a single person, handing an agent a goal and a list of IP addresses and getting nation-state reach in return. AI does not create the vulnerability. It collapses the time and skill needed to exploit it.

There is a second, quieter risk inside our own walls. Organizations are handing their own agents goals like “reduce energy spend” and “optimize grid load,” along with access to building management systems, procurement platforms, and operational technology. With data centers projected to consume a fifth of U.S. electricity by 2035, the pressure to shed load elsewhere will only grow. An agent does not need malice to cause harm. It needs a poorly bounded goal, broad access, and nobody watching. These incidents show how little stands between a remote command and spoiled food, or a compromised blood supply.

And we may not be able to tell the difference. A human attacker, an AI-orchestrated campaign, and a misconfigured internal agent can produce nearly identical outcomes: no note, no manifesto, food spoiling in the dark. Attribution was already hard. Agents make it harder.

We have spent two decades asking whether someone would attack critical infrastructure. The better question for the next decade is whether SOMETHING would, and whether we would even recognize the difference.

The Question Underneath the Question

“What if” is not a prediction, but it forces us to look at a quiet failure in a grocery store freezer and see the shape of something much larger, and much more frightening.

The tools we are building are getting more capable every day. The infrastructure they touch was never designed to stop them. The freezers were just the first place it showed up.

So ask the question. What if this was autonomous agents? How would we respond if we confirmed it was an autonomous attack? Would we slow the development of AI? Would we pull the plug?

Until we can answer that, here is what every leader who owns a building, a plant, or a store can do this week. Find every controller that touches the internet. Change every default password. Put your operational technology on its own network, away from email and the front office. And decide, in writing, who is allowed to hand an AI agent the keys to anything that keeps people fed, warm, or safe. The freezers were a warning. We should treat them like one.