AI models may soon be able to copy themselves across internet-connected devices without needing powerful datacentres. In a new study, Anthropic warned that smaller AI models are getting closer to carrying out this task without human help. These models could spread across inexpensive devices such as Android TV boxes, smart ovens and other connected home appliances. Once spread across a network, an AI worm could be difficult to shut down because it would not depend on a single central server. Anthropic said this threat is not yet fully realised, but advances in AI may bring it closer by the end of 2026.
The researchers tested whether AI models can set themselves up on unfamiliar hardware and spread between machines. Unlike traditional malware, which often follows fixed instructions, an AI-powered worm can adapt to different devices and work out how to run on them.
Also read: Meta unveils Muse Wonder AI models for agentic tasks: Here is what they can do
To demonstrate the computing power available in everyday devices, Anthropic ran the Gemma4-26B-A4B model across 15 inexpensive Android TV boxes. Each box cost around $55 and handled part of the model. Together, the devices ran the AI model at 0.39 tokens per second. The experiment showed that several low-cost devices can work together to run a capable AI model.
Anthropic also tested whether an AI model can spread across a group of virtual Android devices. In a controlled experiment involving eight machines, GLM-5.2 managed to reach all eight devices in four out of five trials.
Also read: Anthropic tightens AI safeguards after Claude sends fake tip to police: Here is what happened
The researchers found that models with 30 billion parameters or fewer can perform many individual steps needed for replication, but often struggle to complete the entire process reliably. Anthropic said that if smaller models continue improving at their historical rate, they could reach self-replicating capability by the end of this year.
The spread of such a worm may also be slow. Anthropic estimated that a small model could take around one to two days to replicate on low-cost hardware.
“We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain (and passing on their adaptations to future copies of themselves),” the company stressed.
To reduce the risk, Anthropic recommended stronger security for internet-connected devices and better network monitoring to identify unusual activity.