In recent weeks, many European supermarkets have discovered that a cyberattack can turn a refrigerator into a source of heat. Modern refrigeration systems use an automatic defrost cycle that periodically suspends cooling and heats certain components to remove ice from the coils, ensuring that the system operates properly. This process can also be controlled remotely. On the night of October 11–12, hundreds of French supermarkets were targeted by cyberattacks that manipulated refrigeration controls, extending defrost cycles and causing substantial food losses. At one Intermarché store in Gradignan, the owner reported that the temperature in a cold room had reached 50°C. On November 2, Reuters documented a new wave of incidents affecting supermarkets in Sweden, Belgium, and other European countries, including another reported case of freezers reaching 50°C in Stockholm. Refrigeration incidents were also reported in Ireland, while Italian authorities confirmed that a limited number of stores had been affected.
In a previous article published on HackerNoon, titled Houston, We Have a Problem: Artificial Intelligence Is Becoming Harder to Control, we had already pointed to a particularly dangerous combination. Over the years, we have connected a growing number of systems to the Internet, often prioritizing convenience and cost savings over security. Controlling equipment remotely avoids costly on-site visits and allows for quick intervention, but it also exposes the system to potential attacks. This growing connectivity is not always matched by adequate security measures: some systems may run outdated software, use default passwords that have never been changed, rely on devices that manufacturers no longer update, or expose services directly to the Internet without sufficient safeguards. And hospitals, transportation, power grids, and other essential services now depend on these infrastructures. Meanwhile, we are developing increasingly powerful AI systems capable of writing code, using tools, and identifying vulnerabilities. So-called AI agent swarms are also emerging: groups of autonomous agents that divide tasks, work in parallel, share information, and adapt their strategies. If used maliciously, these technologies could enable attackers to identify and exploit vulnerabilities at speeds and on a scale that would have been unthinkable until recently. And this is where the paradox lies: as the tools for attacking grow ever more powerful, our society depends on increasingly complex digital infrastructures, often built up in a piecemeal fashion, tightly interconnected, and running software that may contain vulnerabilities.
The supermarket incidents bring an increasingly important question back into focus: if artificial intelligence is becoming such a powerful tool in the hands of attackers, why should it not be equally powerful for those defending against them? The answer lies in a fundamental feature of cybersecurity that AI may amplify: the asymmetry between attack and defense. An attacker may need to find just one vulnerability, while defenders must protect servers, applications, devices, and external connections all at once. Attackers can choose where and when to strike; defenders must be prepared for threats coming from anywhere in their infrastructure. Imagine what might happen if a hacker tasked a swarm of AI agents with autonomously exploring an infrastructure and identifying its weakest points. There is no evidence that autonomous AI agents played a role in the recent supermarket attacks, which illustrate the risks of insecure connectivity rather than AI-driven cyberattacks. The scenario is nevertheless worth considering because increasingly autonomous AI systems could make identifying and exploiting similar weaknesses faster and more scalable.
Using AI for defense, however, is far more complicated. An experiment published by Corma in August 2026 offers an interesting example. The company tested four frontier AI models in 241 attack-and-defense engagements within a controlled enterprise network. The attacking agents attempted to install persistent backdoors, while defending agents were subsequently tasked with identifying them. According to Corma, attackers established verified, reboot-surviving backdoors in 85% of engagements, whereas defenders correctly identified the planted artifacts in 22% of cases involving a verified persistent backdoor. These figures measure different outcomes and come from a specific experimental setup, rather than demonstrating a universal performance gap between offensive and defensive AI. Nevertheless, they highlight an important challenge: successfully exploiting a weakness and reliably detecting its exploitation are fundamentally different tasks.
Detecting an attack is not simply a matter of spotting unusual behavior: an authorized technician and a hacker who has stolen the technician’s credentials could issue identical commands. Moreover, attackers can afford numerous failed attempts, while defenders cannot keep shutting down essential services at every sign of suspicion. An AI system that mistakenly shut down a refrigeration unit could cause damage similar to what it was trying to prevent. Finally, even the most advanced system can only recognize threats whose signals it can observe: if those signals are not collected or interpreted correctly, the attack may go unnoticed.
The challenges of defense become even more apparent when cyberattacks have consequences in the physical world. Cyber tools have also become genuine offensive weapons in contemporary conflicts. Missiles and explosives are not necessarily needed to strike essential services: compromising power grids, hospitals, transportation, communications, or food distribution systems can cause disruption and economic losses and, in the most serious cases, put people’s lives at risk, indirectly or even directly. Of course, not every cyberattack is a military operation, but the digital domain is now an integral part of conflicts. And AI agent swarms, in some respects, resemble another technology that is changing how wars are fought: drones. Although they are very different tools, they share one important feature: they can multiply threats, putting defenders under pressure.
The war in Ukraine has shown how relatively inexpensive drones can be deployed in large numbers, forcing defenders to spend far more to intercept them. A sophisticated interceptor missile can cost many times more than the drone it is meant to shoot down, and when attacks multiply, the economic sustainability of defense becomes a problem in its own right. That is why Ukraine also uses surprisingly simple solutions, such as protective nets installed over roads to hinder drone attacks. According to Ukraine's Ministry of Defence, more than 2,200 kilometres of frontline roads had been equipped with anti-drone protection by November 2026. It is almost a paradox: in a war dominated by highly advanced technology, relatively simple physical barriers are proving useful again. The same lesson applies to cybersecurity: the best response to a sophisticated threat is not always an even more sophisticated technology. AI can help, but we also need updated systems, segmented networks, and, where possible, a decision not to connect devices to the Internet when they do not really need it.
This leads to another question: are we sure that a more digitized society is necessarily a safer one? Some countries have long recognized the importance of maintaining alternatives to digital services. In Sweden and the Netherlands, for example, authorities explicitly recommend keeping cash available in case electronic payments are disrupted. The Swedish central bank suggests around SEK 1,000 per adult, while the Dutch central bank recommends approximately €70 per adult to prepare for a possible 72-hour disruption. In a very different context, during its June 2025 conflict with Israel, Iran severely restricted Internet access, with authorities citing concerns about the alleged misuse of communication networks for military purposes. The restrictions, which culminated in a near-total shutdown, also severely disrupted ordinary citizens' communications. These are very different situations, but they help illustrate an often-overlooked point: depending entirely on digital, interconnected technologies can, in certain circumstances, become a weakness. This does not mean going back to the past or giving up innovation, but preventing a malfunction or cyberattack from paralyzing essential services.
The issue goes far beyond protecting individual systems. Electronic payments, food distribution, transportation, energy, and communications are now closely interconnected. As the supermarket incidents show, even a limited attack can have tangible consequences for everyday life. But what would happen if a disruption affected several essential services and lasted for days? As early as 2025, the European Union adopted a strategy calling for guidelines to help citizens remain self-sufficient for at least 72 hours in an emergency. In the United Kingdom, people are advised to keep supplies of water, food, cash, and even battery-powered radios, useful when the Internet and other services are unavailable. These are general precautions intended for blackouts, natural disasters, and other emergencies as well, not just cyberattacks.
At first glance, these recommendations might seem overly cautious. But when I still read about default passwords that have never been changed, services accessible directly from the Internet without adequate protection, and machines that work perfectly well but run software their manufacturers have stopped updating, I grow concerned. The old IT mantra, “If it ain’t broke, don’t fix it”, is becoming dangerous: a system can continue to work flawlessly while remaining vulnerable to attack. And my concern grows when I think about the complexity of IT infrastructures, built up over time by layering software developed by different people and companies, often using different technologies, and connecting an ever-growing number of services. Add to that cameras, sensors, and other devices sometimes purchased without sufficient checks on their security. And as we continue to make everything more connected, we are developing AI systems increasingly capable of finding and exploiting precisely these weaknesses.
All the ingredients for a perfect mess are there. Let's hope for the best.

