Small AI models will soon spread across consumer devices, including smart home devices and lower-cost Android TV devices.
This is according to research by Anthropic. The company says AI-driven “worms” could be a reality as early as the end of this year.
In a series of experiments, Anthropic investigated whether AI models could replicate themselves without relying on conventional data centers. The researchers concluded that models with fewer than 30 million parameters can already perform the individual steps needed for autonomous replication. However, these models are not yet reliable enough to do so consistently.
In one experiment, researchers deployed Google’s Gemma4-26B-A4B model on fifteen Android TV boxes, each costing only $55. The experiment showed that relatively powerful AI models can run on distributed consumer hardware and therefore don’t require specialized servers.
A new generation of botnets
Unlike conventional malware, which typically relies on predefined instructions, AI-driven worms could adapt their behavior to unknown hardware and software environments. This would let the malware spread across different types of devices.
Anthropic also tested the larger GLM-5.2 model, which successfully replicated itself in a simplified experimental environment. The researchers note, however, that the results do not show the models can autonomously spread across networks in practice.
The researchers note that this distinction is important. In the experiments, the researchers simplified key aspects of device takeover, such as gaining access to the devices. Bypassing security measures likely remains a significant hurdle for the “worms.” Nevertheless, Anthropic expects smaller AI models to have the necessary capabilities within a few months.
Difficult detection
In practice, these systems are likely to spread slowly. However, their distributed nature could make them difficult to detect and remove with conventional methods.
According to Anthropic, the study’s findings raise concerns about a scenario in which AI-driven malware is no longer dependent on centralized infrastructure. Instead, compromised consumer devices could collectively provide the computing power that autonomous AI systems need to spread.