# Anthropic: small AI models may soon copy themselves across smart TVs and ovens

URL: https://ai-pulse.aglaganova.com/blog/anthropic-small-ai-models-may-soon-copy-themselves-across-smart-tvs-and-ovens
Published: 2026-10-14
Author: Roman Aglaganov, Full-stack developer
Category: Research
Region: Global
Tags: safety, open-weights, agents

> Anthropic’s Frontier Red Team ran a 26-billion-parameter model on fifteen $55 TV boxes, and found one open model that can already spread itself across a test network. The worrying part is the timeline.

![Illustration: fifteen small TV boxes in a grid, with a red chain linking the ones that have turned black](https://ai-pulse.aglaganova.com/api/media/file/replicate-1.png)

A computer worm copies itself from machine to machine without anyone’s help. Security researchers have long worried that AI could make worms smarter. A new study from Anthropic’s Frontier Red Team, published on 12 October, asks a more specific question: could a small AI model carry its own “brain” with it, running on the cheap, poorly secured gadgets in our homes?

## The experiments

**Cheap hardware is enough:** the researchers bought fifteen off-brand Android TV boxes for about $55 each, networked them, and ran Google’s Gemma4-26B-A4B model split across them, at 0.39 tokens per second. Slow, but working.

**Setting up is the hard part:** they tested whether open models could get an AI model running on unfamiliar hardware. A 9-billion-parameter model managed it on 8 of the 9 hardware types tried, from Jetson Nano boards to Macs.

**One model can do the whole loop:** on a sealed network of eight emulated Android machines, the team checked whether a model could find a new device, copy itself over, start running and repeat. Only GLM-5.2 managed it, reaching all eight nodes in four of five trials.

The authors are careful about limits. The hacking step was made trivial (an unprotected port), so the result is an upper bound. And GLM-5.2 is far too big to run on a TV box itself.

## Why the timeline matters

The concern is the trend. Plotting small open models against a general capability index, the authors estimate that a model small enough for such devices could reach GLM-5.2’s level by the end of 2026. They also found the models can learn as they go: copies that passed notes to their successors got better at setup, rising from 45% success in the first generation to between 61% and 84% later.

Such a worm would be slow, with copies taking a day or more, the authors estimate. But it would have no central off switch, and unlike ordinary malware it could work out unfamiliar devices on its own. The old advice now matters more than ever: change default passwords, update devices and keep cheap gadgets off important networks.

## Sources

- [Anthropic: Small AI models may soon be able to self-replicate on IoT devices](https://www.anthropic.com/research/self-replicating-ai)