Anthropic Warns AI Worms Could Spread Across Connected Smart Devices

Anthropic has warned that small AI models could soon replicate themselves across internet-connected devices, potentially creating distributed networks that are difficult to shut down.

In a research paper published on October 12, researchers from Anthropic’s Frontier Red Team said models could be months away from autonomously copying themselves onto devices such as Android TV boxes and smart appliances. Unlike conventional malware, these AI “worms” could adapt to unfamiliar hardware and set up models to run on compromised devices.

The researchers tested whether models could prepare unfamiliar hardware to run AI workloads. A 9-billion-parameter model successfully set up inference on eight of nine hardware types. In a separate experiment, researchers used the GLM-5.2 model to propagate copies of another model across a firewalled network of eight cloud-based Android emulators. It reached all eight machines in four out of five trials.

The experiments did not function as a fully autonomous worm spreading across real-world consumer devices. Anthropic’s timeline is a projection based on how quickly small models have improved. The researchers estimate that models with these capabilities could emerge by the end of 2026 if that progress continues.

A successful self-replicating system could use compromised devices' computing power to run AI models without relying on a single data center. Because the network would have no central point to disable, containing it could require identifying and cleaning each infected device. The company thus recommends stronger security for connected devices and further work to prepare for this potential threat.