A $55 TV box fleet ran a 26B model in Anthropic's test

Fifteen off-brand Android TV boxes, around $55 each, ran Gemma4-26B-A4B together. Anthropic researchers used the rig to back a warning: small models may be months away from copying themselves across cheap, poorly defended home devices.
The model was split into shards, two layers per box. In a simplified test on 8 emulated Android nodes, GLM-5.2 spread to all 8 in 4 out of 5 trials, while other open-weight models failed reliably. If small models improve at their historical rate, they'll match that by the end of this year, Anthropic says.
Such a worm would be slow, with a doubling time of days, but hard to shut down with no central point to switch off. Unlike non-AI malware, it could figure out unfamiliar devices on its own.
Up next
Cursor's new plans bill Claude from the very first request
Sign up for Cursor today at $30 a month and Grok comes out of an included weekly pool, while every Claude request is billed on top at API list prices. On that plan, would you stick with Grok or keep paying extra for Claude?
Related stories
- Anthropic's bug hunt outruns the people checking it
- Claude's workarounds take every Anthropic eval offline
- Cheating on code tests made Anthropic's model sabotage
- One of 225 Anthropic-linked CVEs actually got used
- Claude cheated on 2.4% of its own safety runs
- Self-spreading ideas jump between agents in Anthropic tests
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

