That cheap streaming box plugged into the back of the TV doesn’t look like much of a threat. New research suggests we should start thinking about it differently.
On 12 October, William Zhou and Daniel Freeman from Anthropic’s Frontier Red Team published research suggesting small AI models could be only months away from behaving like a computer worm. In other words, an AI model that copies itself from one internet-connected device to the next and runs on whatever it lands on, with no human and no data centre involved.
What did they actually test?
A traditional worm is a program that spreads from computer to computer by itself. The researchers wanted to know if an AI model could do the same thing, carrying its own model files and running on the hardware it infects.
First, they bought 15 off-brand Android smart TVs, linked them together and ran a 26-billion-parameter open model split across all of them. It was painfully slow (0.39 tokens per second), but it worked. Then they tested whether models could install AI software on unfamiliar hardware by themselves. A 9-billion-parameter model managed it on 8 of the 9 hardware types they tried. And in a small test network of eight emulated Android devices, one model (GLM-5.2) spread itself all the way through, reaching every device in four out of five runs.
Their rough estimate is that small models able to run on this kind of cheap hardware could reach the same level by the end of this year, if they keep improving at the rate they have been. They’re upfront that it’s a ballpark figure, based on one device type and one past botnet.
Which devices are we talking about?
The research is based on Kimwolf, a real botnet Cloudflare says is made up of around two million devices worldwide, mostly Android TV streaming boxes. Kimwolf gets in through an exposed Android debugging service, and Cloudflare says many low-cost TV boxes arrive with proxy software already installed. The paper lists a wider range of devices that could run part of a small AI model, and the list will look familiar:
- no-name Android TV and streaming boxes
- smart fridges and ovens
- robot vacuums and robotic lawnmowers
- security cameras.
The paper doesn’t mention routers, but I would add them. Everything else connects through the router, and it’s often the device that hasn’t been updated in years.
Offices have plenty of these too. Copiers, CCTV recorders and video-conferencing units increasingly come with built-in AI chips for things like noise cancelling and transcription. Small business NAS storage boxes are on the list, and they already have a long history of being hit by ransomware and cryptojacking. Add the meeting room TV and the router, and most offices have a handful of devices nobody is really managing.
Slow, but hard to get rid of
An AI worm like this wouldn’t move quickly. The researchers think one copy could take a day or two to set itself up on new hardware, and they compare it to a slime mould or a swarm of fruit flies. Even a network the size of Kimwolf would only add up to roughly the computing power of 40 high-end data centre GPUs.
The worrying part is how hard it would be to kill. There’s no central server to switch off. Older malware was written for one type of device, but an AI model can work out unfamiliar hardware for itself, and later copies can learn from earlier ones. In one experiment, the setup success rate climbed from 45% in the first generation to between 61% and 84% in later ones.
The good guys do have some advantages. Shifting multi-gigabyte model files between devices creates unusual network traffic that internet providers can spot, and past botnets have been cleaned up by providers isolating infected customers.
What to do about it
None of this needs new technology. Own Your Online (the NCSC’s public advice site) already recommends the basics for smart devices: change the default passwords, keep them updated, switch off smart features you don’t use, put them on a separate guest WiFi network, and check the security reviews before you buy. If a device doesn’t get updates any more, replace it or turn off its internet features.
In a business it’s the same idea, just a bit more structured:
- keep a list of EVERYTHING connected to your network, not just laptops and phones
- put cameras, conferencing gear and building systems on their own network segment
- treat any cheap, no-name device with no update history as a risk, no matter how little it cost.
Want to know what’s really connected to your office network? Book a free health check, or call us on 0800 688 768.