Anthropic warns that small AI models may soon be capable of self-replicating on IoT devices, potentially forming autonomous worm botnets without a central shutdown point.
Testing across open-weights models explores propagation risks on connected hardware.
According to the company, such programs could form autonomous networks that spread from device to device without a central point to easily shut them down. Previous work has shown systems scanning networks, exploiting security vulnerabilities and copying their files to data centers.
Such a program could carry its files with it or re-download them from a public platform, so that it can run directly on compromised hardware. Malicious actors could build decentralized networks of poorly protected smart devices. These programs spread by compromising the computers they run on. Creating hidden installations in large data centers is considered a more advanced category, while other methods rely on stealing digital connection keys from artificial intelligence services to run copies of themselves.
Anthropic tested 11 open-source models from companies including Alibaba, DeepSeek, Google, Meta, MoonShot AI, OpenAI, and Zhipu AI. It found that one particular model, GLM-5.2, was the only one that was able to reliably transmit another program to test nodes running an Android emulator in four out of five tests.
A network of cheap Android TV devices could form a slow, decentralized system that works together to run computational tasks. The company says it doesn't yet have a fully functional program of this kind, but systems are becoming increasingly capable of completing individual steps in the deployment process.
Top comments (0)