Anthropic researchers are warning about a potential scenario in which small AI models begin spreading between inadequately secured devices and using their computing resources without the owners’ knowledge. Such a network could include smart TVs, refrigerators, robot vacuum cleaners, and other internet-connected appliances. However, this is a projected risk rather than evidence of a confirmed epidemic of AI malware: the experiments described so far do not prove that such systems are already infecting household devices on a large scale.

How an AI worm differs from a conventional virus

Classic computer worms are malicious programs capable of copying themselves and spreading across a network on their own. In the scenario described by Anthropic, a malicious system could transfer to other devices not only software code but also the AI model itself, and then use it to perform tasks on the new hardware.

The researchers also describe the possible spread of such systems according to the principle of “digital mold”: small models occupy the available computing resources of many devices and gradually form a distributed network.

The potential problem is that household appliances often have limited security protections and are rarely perceived by their owners as полноценные компьютеры, requiring regular security oversight. At the same time, even low-powered devices may have enough resources to run compact models.

Experiment with low-cost TV set-top boxes

As part of the study, the team tested whether small AI models could run on affordable hardware. In particular, Google DeepMind’s Gemma model was successfully run on 15 low-cost Android TV set-top boxes from an unknown brand, priced at about $55 each.

The researchers also studied models from Chinese developers, including Alibaba’s Qwen and a GLM-family model from Z.ai. These experiments helped assess the capabilities of small models and the availability of computing resources outside server centers.

However, running a model on a cheap set-top box does not in itself mean that the device was infected or that the model was autonomously replicating across the network. This is an important distinction: the experiment demonstrates the technical possibility of using affordable hardware for computing, whereas a full-fledged AI worm would additionally need to be able to penetrate other devices and spread its model.

Why such a network would be hard to shut down

If such a system really can spread between devices, it will be harder to control than a program running on a single server. Computation would be distributed across many devices belonging to different people and organizations.

Some of the equipment may operate for years without security updates. Owners might not even suspect that their appliances are performing unrelated tasks. In such a case, attackers could use the combined computing power of infected devices, although the exact scale and practical efficiency of such a network have not yet been established.

Anthropic views this scenario as one of the possible risks of further development of autonomous AI systems. The researchers warn that in the future, software agents may run on many different types of hardware, including devices that are not under direct human control.

What can be done right now

The main protective measure is to prevent connected devices from being left without updates and basic cybersecurity safeguards. It is important for manufacturers to provide vulnerability patches, restrict application access to system resources, and ensure secure software updates throughout the device’s entire service life.

Users should install firmware updates, change default passwords, disable unnecessary network features, and, where possible, place smart home devices on a separate Wi-Fi network. This does not guarantee protection from every threat, but it reduces the opportunities for unauthorized access.

Anthropic emphasizes that preparing for future threats should include not only preventing breaches, but also the ability to detect attacks and limit their consequences if they cannot be fully ruled out.

For now, “AI mold” remains a possible threat scenario rather than an established fact of mass infection of household appliances. But the study draws attention to an important issue: as compact AI models become more widespread, computing resources may appear almost everywhere — and with them, new ways to abuse devices that were not previously considered attractive targets for sophisticated attacks.