A warning from the lab
The research, led by William Zhou and Daniel Freeman of Anthropic’s frontier red team, asked a simple question: could an AI model find vulnerable devices, copy itself onto them and keep spreading without human help? The team tested multiple models developed in the United States and China.
The results showed that no current model is both powerful and small enough to replicate reliably on a smart TV. But the gap is closing fast, and based on current trends the researchers expect such a model to appear before the end of 2026.
“Home and small business IoT devices might be the first target for such attacks.” — William Zhou and Daniel Freeman
Inside the experiments
The team broke self-replication into four stages: finding targets, gaining execution, setting up AI inference on the new device and propagating from that host. Some notable results:
In a firewalled eight-node cloud environment, GLM-5.2 spread to all eight nodes in four of five trials, and to seven of eight in the fifth.
















