Anthropic researchers ran a 26-billion-parameter AI model across 15 inexpensive Android TV boxes on October 12. The test did not demonstrate a small model autonomously infecting devices and spreading. It showed that AI inference could run across low-cost hardware, one step in a possible malware chain.

That distinction matters. A model that can configure devices is not the same as a computer worm that finds targets, breaks in, installs itself and repeats the process.

What the TV-box test showed

Anthropic’s team split the Gemma4-26B-A4B model across the 15 off-brand Android TV boxes. Together, they generated 0.39 tokens per second, a slow rate that shows the practical limits of pooling cheap hardware.

In a separate test, a 9-billion-parameter model got inference running on eight of nine hardware types. It showed the model could adapt its setup to different devices. It did not show those devices being infected by the model.

  • 15 Android TV boxes ran the distributed model.
  • The boxes generated 0.39 tokens per second.
  • A separate model ran on eight of nine hardware types.

The missing link: a complete infection

Anthropic said smaller models could complete many individual replication steps but made mistakes when combining them. Its report also describes successful end-to-end results in simplified tests with GLM-5.2 and stronger models.

Ars Technica noted that the research did not show small models independently carrying out every step of an infection. It also reported that GLM-5.2 coordinated part of the process.

There was another important limit. Ars Technica reported that the Android devices exposed an unauthenticated Android Debug Bridge, a debugging interface that should not be open on production devices. That makes the setup a poor match for a typical, well-secured home device.

A warning about what could come next

Anthropic’s concern is a forecast, not a report of an active AI worm. The company argues that improving small models may eventually make it easier to run AI on compromised consumer devices without relying on a data centre.

The experiment measured part of that possibility, not a complete, autonomous infection.

Low-cost devices can host distributed AI inference. This test did not prove that small models can independently turn them into a self-spreading network.