Researchers at Anthropic ran a model with 26 billion parameters on 15 $55 set-top boxes and measured the ability of small models to propagate independently from machine to machine.
A team from Anthropic has published a study on the autonomous replication of AI models, i.e. their ability to copy and restart on other machines without human intervention, like a computer worm whose model weights would form the payload. Its conclusion: small models could be only a few months away from being able to do so on consumer connected objects, such as smart ovens or unbranded, poorly protected Android TV boxes.
A botnet of connected objects is enough
To demonstrate this, the researchers bought and networked 15 TV boxes for about $55 each, with 4 GB of memory, and ran Gemma4-26B-A4B on them by distributing the layers of the model between the boxes, at 0.39 tokens per second. In their tests, a 9 billion parameter model managed to get inference up on 8 of the 9 types of hardware tried, from Jetson Nano to Macs. Models with 30 billion parameters or less already pass most of the isolated steps, but make mistakes by chaining them together; in a simplified environment of eight emulated Android nodes, GLM-5.2 and more powerful models complete the full replication cycle.
If small models continue to grow at the rate observed so far, they would reach this level by the end of the year, according to Anthropic. Such a replicator would be difficult to turn off due to the lack of a central point, even if its propagation would remain slow (doubling in a few days), and it could, unlike conventional malware, adapt on its own to unknown devices. Its processes do not necessarily require administrator rights. The post concludes with some defense against this type of threat.
Read the full article on the original source
Go to the source →