Small AI models are getting close to copying themselves from one cheap smart device to the next with nobody involved. The models that can do the whole job are still too big to fit on a TV box, but the gap is closing, and the first home for a self-copying AI is likely to be the cheapest, least-watched hardware in a house or an office.
Odd devices were safe because attacking them took effort
Cheap devices have not been safe because they are well built. They have been safe because each make of TV box, camera or oven runs slightly different software, and old-style malware has to be written for each one. Mirai spread so far because one trick, a list of factory passwords, fit a huge number of near-identical devices. The odd ones were not worth a person's time.
A model does not have that problem. Setting up an AI model on a new machine means choosing software that fits its chip and operating system, finding the right versions, and fixing whatever breaks, and each step depends on the one before. In tests of that whole job, a 9-billion-parameter model managed it on 8 of 9 kinds of hardware, from a small Nvidia board to a Mac.
The test devices had no password at all, so real success will be lower. But breaking in is the part that is already cheap: a free model can already build working exploits, and Mirai got in with a password list.
It would be slow, and nobody would be running it
The copies would not be fast. Fifteen cheap TV boxes working together ran a 26-billion-parameter model at about one word every three seconds, and getting one new copy running takes a day or two. Even a fleet as large as the two million boxes that the Kimwolf botnet took over would add up to roughly 40 datacenter chips' worth of computing.
What makes it hard to deal with is that no one is in charge of it. There is no server to seize and no account to close, so each infected device has to be cleaned one at a time. The copies can also pass notes to the next one about what worked. In a test, the share of copies that finished setup rose from 45% in the first generation to between 61% and 84% in later ones. So each kind of device gets easier after the first.
As for timing, the model that managed the whole chain is too big for a TV box. Small models have improved at a steady rate, and extending that line puts one at the same level by the end of 2026. That is a rough estimate from a general score, but the direction is clear.
What to do about your own gadgets
The defenses are the ones that worked against Mirai. When an internet provider told customers their devices were infected and limited them to cleanup pages, 92% of infections were gone within 14 days, against 74% when customers were not told. So if your provider sends a notice, act on it the same day.
Beyond that, change every factory password, and put TV boxes, cameras and conference units on a guest network, away from the laptops that hold work files. Kimwolf already scans the local network of each box it takes over, looking for more devices to infect, so a box on the main network is a stepping stone. Replace off-brand boxes that never get updates. In the UK, default passwords are no longer allowed on new smart devices, but you should not assume the same rule protects you elsewhere.
The same ability can work for the person looking after these things. The skill that makes a worm dangerous, working out a strange machine without a manual, is also what a shop owner needs to look after a dozen gadgets. An assistant with that skill could list everything on the router, point out which ones still use factory passwords, and walk through changing them.