Most people wouldn’t think of the cheap streaming box plugged into the back of the TV as a cyber risk. But, thanks to new research from Anthropic on AI worms and smart devices, it seems that we should.
On 12 October, two researchers from Anthropic’s Frontier Red Team, William Zhou and Daniel Freeman, published a post arguing that small AI models may be only months away from behaving like a computer worm. That means a model that copies itself from one internet-connected device to the next and runs on whatever hardware it lands on, with no human involved and no data centre needed.
What Anthropic tested
A traditional worm is a program that spreads between computers on its own. The researchers wanted to know whether an AI model could do the same, carrying its own model files with it and running on whatever it infects.
To show that cheap hardware has enough combined computing power, they bought 15 off-brand Android Smart-TVs, networked them together and ran a 26-billion-parameter open model split across all 15. It was slow, at 0.39 tokens per second, but it worked.
They then tested whether models could set up AI software on unfamiliar hardware by themselves. A 9-billion-parameter model managed it on 8 of the 9 hardware types tested. In a simplified test network of eight emulated Android devices, one model, GLM-5.2, spread itself end to end and reached all eight nodes in four of five trials. Their estimate is that small models able to fit on this kind of hardware could reach the same level by the end of this year, if they keep improving at their historical rate. They’re clear that it’s a rough estimate, based on one type of device and one past botnet.
The devices in question
The research is modelled on Kimwolf, a real botnet that Cloudflare says is made up of about two million devices worldwide, mostly Android TV streaming boxes. Kimwolf spreads by exploiting an exposed Android debugging service, and Cloudflare notes that many low-cost TV boxes arrive with proxy software already installed.
Anthropic’s paper names a wider set of devices that could run part of a small AI model. In most New Zealand homes the list will look familiar:
- Off-brand Android TV and streaming boxes
- Smart fridges and ovens
- Robot vacuum cleaners and robotic lawnmowers
- Security cameras
You could assume other equipment is also vulnerable, included your home router. The paper doesn’t name it, but everything else connects through it, and it’s often a device that hasnt been updated in years.
Many of the same items are also repsent in offices. That includes copiers, security cameras and video-conferencing units, which increasingly ship with built-in AI chips for things like noise cancellation and transcription. It also mentions small business NAS storage boxes, which have a history of being hit by ransomware and cryptojacking. Add the meeting-room TV, the CCTV recorder in the comms cupboard and the router, and most offices have a handful of devices sitting outside normal IT management.
Slow, but hard to kill
An AI worm like this wouldn’t be fast. The researchers estimate a single copy could take one to two days to set itself up on new hardware, and compare it to a slime mould or a swarm of fruit flies. Even a fleet the size of Kimwolf would add up to a modest amount of computing power, roughly 40 high-end data centre GPUs by their reckoning.
What makes it worrying is persistence. With no central server to switch off, a self-copying worm would be hard to stamp out. Older malware was written for one type of device, whereas an AI model could work out unfamiliar hardware for itself, and later copies could inherit notes from earlier ones. In one experiment, setup success climbed from 45% in the first generation to between 61% and 84% in later generations.
Defenders do have some advantages. Moving multi-gigabyte model files between devices creates unusual network traffic that internet providers can spot, and past botnets have been cleaned up by providers quarantining infected customers.
Tightening up
None of this needs new technology. Own Your Online, the NCSC’s public advice site, already recommends the basics for smart devices: change default passwords, keep software updated, turn off smart features you don’t need, put smart devices on a separate guest WiFi network, and check security reviews before you buy. If a device no longer gets updates, it suggests replacing it or switching off its internet features.
In a business, the same thinking applies with a bit more structure. Keep an inventory of everything connected to the network, not just laptops and phones. Put cameras, conferencing kit and building systems on their own network segment. Most importantly, we now need to treat a no-name device with no update history as a risk, however cheap it was.