AI models can replicate themselves to smart devices

Small AI models may be able to replicate themselves onto smart devices such as smart TVs, ovens, and refrigerators within a few months. This is reported by Anthropic, the company behind the AI assistant Claude. If such "AI worms" circulate, they could be difficult to stop, as they do not possess a central off switch. This is evident from research published on October 12, 2026, in which AI models were able to independently find and take over other computers within a closed network.

In Anthropic's experiments, researchers found that a small AI model with 9 billion parameters successfully ran on 8 of the 9 different devices tested, ranging from mini-computers to Macs. It appears that these types of models can install themselves and operate independently on unknown hardware. The expectation is that such small models could reach their full potential by the end of this year if they continue to improve at the recent pace.

To illustrate the computing power of common devices, the researchers used fifteen cheap Android TV boxes, which collectively ran a single AI model. This demonstrates that inexpensive and widely praised products like these present an attractive target for attackers. The Kimwolf botnet, for example, has taken over more than two million such devices. A network of this nature would be capable of running thousands of copies of an AI model simultaneously, further increasing the threat of an AI worm.

While the spread of an AI worm is slow, the model learns from each step it takes. The success rate for successfully installing the AI on new devices has risen from 45% to 61-84% as more copies are created. In response, Anthropic points to research from TU Delft, which indicates how infected devices can be quarantined and cleaned, resulting in a significant reduction in infections. There is a call for critical scrutiny of cheap, unbranded devices and the importance of adequately updating these products.

Read the full article from AI Insider.