Across poorly secured consumer devices, small AI models might be able to copy themselves within just a few months. That warning comes from a new study authored by two Anthropic researchers, William Zhou and Daniel Freeman.
The two set out to test whether such a model could propagate over the sort of hardware people casually leave sitting on a home network. Their scenario is an AI worm that bundles model weights inside its malware payload and then runs inference straight on whatever device it manages to compromise. Adaptability is what sets it apart from ordinary malware: a conventional worm sticks to a fixed script, whereas a model can probe unfamiliar hardware and figure out on its own how to configure itself to run there.
Fifteen off-the-shelf Android TV boxes served as the test bed, together running the Gemma4-26B-A4B model. RAM on each was 4GB and the price tag ran about $55, yet no single unit held the whole model — layers were split among the machines. Performance was slow, however it worked.
Four phases made up the replication chain. A network gets scanned for vulnerable devices first; access is then obtained; the AI environment is installed on the target; and from there the process spreads. Because the intrusion step was deliberately omitted from testing, the demonstration never reaches a complete end-to-end attack.
Chief executive Dario Amodei has placed the possible damages in the hundreds of billions of dollars, a projection that presumes widespread insecure IoT hardware alongside scant willingness to fix it.
