AI Worms Could Turn Your Smart Home Into a Botnet, Researchers Warn

Your smart home is a network of small computers: the vacuum that maps your floors, the thermostat that learns your schedule, the TV that streams while connected to the internet. About half of German households now own at least one such device, according to industry data. They talk to each other. They make life easier. And researchers at Anthropic have just published a warning that should make you think twice about how many of them you own.

The threat is not a traditional computer virus. Those have existed for decades—self-replicating programs that exploit known security holes and spread from machine to machine by following a fixed set of instructions. Once you patch the vulnerability, the virus stops spreading. It is a problem with a solution. An AI worm would be something else entirely. Instead of executing a predetermined attack, it would solve problems on the fly. It would look at a new device—one running different software, with different memory constraints, different defenses—and figure out how to infect it. It would learn to bypass security updates rather than depend on a single exploitable flaw. It would adapt. And it would keep adapting as it moved from device to device.

Anthropicresearchers tested whether this was actually possible. They took fifteen cheap Android TV boxes and distributed an AI model across them, with each device doing part of the computational work. This proved that sophisticated AI does not require a massive data center; it can run on the kind of hardware sitting in millions of living rooms. Then they set up eight more simulated Android devices in a controlled environment and unleashed the AI model. In four out of five attempts, it infiltrated all eight target devices. In the fifth attempt, it reached seven. The researchers estimate that an AI worm could establish itself on a previously unknown device in one to two days.

The implications compound when you consider scale. A single infected smart speaker or TV box has limited processing power. But link hundreds or thousands of them together—a botnet made of household devices—and you have a distributed computing network that grows stronger as it spreads. The malware would not depend on a central server that authorities could shut down. It would learn continuously. It could potentially jump between entirely different types of devices, each running different software. Anthropic compares the scenario to a digital slime mold, slowly consuming everything in its path. Even if you disconnect your device from the network or patch a security vulnerability, the decentralized structure would make containment extraordinarily difficult.

The real-world stakes became visible in France in mid-October. Overnight into Monday, attackers compromised refrigeration and freezer systems at numerous supermarkets. Temperatures in the affected units climbed to fifty degrees Celsius. Food spoiled. Some incidents involved fires and equipment damage. France's cybersecurity agency, CERT-FR, initially identified systems made by Danish manufacturer Danfoss as compromised, then expanded the warning to include other manufacturers. The investigation is ongoing, but the pattern is clear: modern supermarkets rely on digital controllers to monitor and regulate refrigeration. Maintenance companies access these systems remotely. Some of the security software was outdated—a gap that attackers apparently exploited.

No evidence has emerged that a self-replicating AI worm was used in the French attacks. But Anthropic's research shows how significantly AI could lower the barrier to entry for such attacks. You would not need sophisticated infrastructure or deep technical knowledge. You would not need to find and exploit a specific vulnerability. An AI worm could do the reconnaissance and adaptation for you. The consequences could be severe: not just spoiled food, but disrupted power grids, compromised medical devices, or industrial systems brought offline. France's cybersecurity agency has advised all companies to update connected devices regularly, install security patches, and disable internet access wherever it is not essential. It is a reasonable response to a threat that is not yet here but is moving closer.