The worry about AI copying itself across the internet has mostly lived in thought experiments. Anthropic has now put a rough date on it. The company said small AI models may be able to copy and spread themselves across connected devices by the end of 2026, if they keep improving at the current rate.
The targets in Anthropic's experiments weren’t racks of servers in a data center. They were Android TV boxes and smart home devices, the cheap gadgets that tend to sit on a shelf for years without anyone checking on them.
Small models can do the steps but not the whole job
The study asked a narrow question. Could an AI system find vulnerable devices, install copies of itself on them and keep spreading without any human intervention whatsoever?
For models with fewer than 30 billion parameters, the answer was mixed. They could complete individual steps of the process. But they struggled to put the entire sequence together.
Larger systems did better. GLM-5.2 and more powerful models achieved full end-to-end self-replication, though only in simplified tests where certain security measures had been removed.
Read the fine print before you panic
That caveat matters. A model that can only finish the job once some of the defenses are gone hasn’t shown it can break into a properly locked-down network. What the results do show is that the gap between “can do each step” and “can do all of them” is getting smaller.
Anthropic’s late-2026 estimate depends on one condition. Small models have to keep improving at their current pace. The company said “may,” and that word is carrying a lot here.
Slow to spread and hard to stop
Even if it happens, Anthropic doesn’t expect a fast outbreak. The company expects any spread to be slow and to depend on widely available computing resources with low security standards.
Containment is where things get uncomfortable. Anthropic warned that without a central “kill switch,” these models would be difficult to contain. Once a copy is running on a device you don’t control, nobody can shut every copy down from one place.
The fix Anthropic is pushing isn’t a new safety feature inside the models. It’s calling for stronger device security across the board, which puts much of the burden on the companies that make the hardware and the people who own it.
So the advice is simple. If an Android TV box or a smart home hub on your network hasn’t had a firmware update since you took it out of the box, it’s exactly the kind of hardware this research has in mind. Update it, or take it off your network.
