What Is an AI Worm? Expert Warns of Risks to Fridges, Water Systems

Scientists at an AI company called Anthropic have released new research warning that a new kind of computer threat could be coming. They call it an 'AI worm' — a program that could copy itself from one device to another without any help from a person. The research says small AI models may be only months away from being able to do many of the steps needed to spread on their own. Experts say this could one day affect not just computers, but also smart home devices, refrigerators, and even systems that treat our drinking water.

A regular computer worm is a program that copies itself from machine to machine without needing a human to click anything. An AI worm would work the same way, but it would use an artificial intelligence model to help it figure out how to get into different devices. This means it could potentially change its methods depending on what kind of device it is trying to infect. That makes it harder to defend against than older types of malware, which were usually built to target just one kind of system.

John Walsh is an expert in cybersecurity who works for a company called IGEL. He spoke to Newsweek and explained why this threat is different from normal hacking. He warned that an AI worm could go after more than just computers and phones. 'The danger is not simply that a computer becomes infected,' Walsh said. 'It is that unauthorized digital access could ultimately influence refrigeration, water treatment, pumping systems or other essential operations.'

Walsh also pointed out that smart home devices — like fridges, ovens, thermostats, and security cameras — could be targets. These devices often have weak security protections and don't get updated often. Because so many of them are connected to the internet, they could be used as stepping stones for a worm to spread further. 'Many have limited security protections, inconsistent software updates and long operating lifespans,' Walsh said, adding that this makes them attractive targets.

Anthropic's researchers ran several experiments to test how close this threat really is. In one test, they spread an AI model called Gemma4-26B-A4B across 15 cheap Android TV boxes, where each box held just a piece of the model. Together, the boxes could run the model — though slowly. This showed that even low-cost, everyday devices could work together to power an AI model.

In another experiment, researchers set up a fake network of eight virtual computers to see if an AI system could spread through them. The AI they tested, called GLM-5.2, reached all eight virtual computers in four out of five tries. However, the researchers made this experiment easier than a real-world attack would be, and they described this test as showing the best possible outcome, not what would typically happen.

Anthropic was clear that it has not shown a real AI worm spreading through actual homes or real devices. The experiments were done under controlled conditions, meaning researchers set things up to make the process easier than it would be in real life. The company said that current AI models can do many of the individual steps needed to spread, but they still make mistakes when trying to put all those steps together. Anthropic wrote that they do not yet have a 'full rogue replicator.'

The bigger concern is what could happen if a worm reached systems that control real-world equipment, like water pumps or food storage units. Walsh explained that just infecting a device is not the same as controlling what it does physically. 'Compromising a device is not the same as gaining authority to control its physical functions,' he said. Getting true control over a refrigerator's temperature or a water treatment process would require extra steps beyond simply infecting the device.

Still, experts say we shouldn't wait for an attack to happen before we start protecting ourselves. Walsh said organizations need to use 'Zero Trust' security, which means no device is automatically trusted just because it's connected to a network. Networks should be split into smaller sections so that if one device is infected, the worm can't easily jump to others. 'Compromising one device must not allow an attacker to move freely across networks or disrupt food supplies, water systems, or other essential services,' he warned.

Anthropic estimates that if small AI models keep improving at the rate they have been, a model powerful enough to pose a real risk could exist by the end of 2026. The company said this is a rough guess based on general trends, not a guaranteed timeline. For now, researchers say the best defense is to build strong security into devices and networks from the very start. The goal is to stop attacks before they happen, not just try to fix things after a device has already been infected.