In one test, researchers connected 15 inexpensive TV boxes and split an open-weight Google DeepMind Gemma model across them. Each box processed only a small number of model layers, allowing the full model to return responses although no individual box could run it alone.
The team also tested propagation in a virtual sandbox. A larger model from Z.ai was able to get a smaller Meta Llama model running on target devices. The larger model, however, was too large for those weak devices and continued to rely on a separate server.
The article therefore says the work did not demonstrate a self-sufficient AI worm that can both spread and run itself on weak smart devices. Security researchers quoted by Scientific American noted that self-replicating malware, botnets, and hijacked computing resources are established threats. The source does not establish that such an autonomous IoT worm exists in the wild, nor that Anthropic's tests produced one.