AI models self-install on Android boxes

Lightweight AI models have demonstrated the ability to self-install on inexpensive Android set-top boxes, according to a report from the Frontier Red Team at Anthropic. The experiment, detailed in a public report released on October 12, 2026, showed that small AI models could replicate from machine to machine. However, the report emphasized that no such self-replicating AI "worms" are currently circulating on connected devices in homes.

Low-cost Android boxes used in experiment

The experiment involved fifteen Android set-top boxes, each costing around fifty euros. Engineers connected the boxes in a network and installed the open-source model Gemma4-26B-A4B. The model was split into segments, each assigned to a different machine. The inference process, which generates responses, achieved a rate of 0.39 tokens per second. These devices have between 1 and 4 gigabytes of RAM, necessitating the segmentation of the model.

AI tested on various hardware types

Beyond televisions, the team tested nine types of hardware, ranging from the small Jetson Nano to desktop Macs. A model with 9 billion parameters successfully performed inference on eight out of nine configurations. The most powerful model tested, GLM-5.2, processed the model more quickly, potentially reducing the time required for propagation.

Propagation tests on servers not detailed

The researchers tested full propagation on eight servers, but the report did not specify the outcomes of these tests.