Tech

AI agents are predicted to take over household appliances

Anthropic warned that by the end of 2026, artificial intelligence agents could learn to self-replicate through smart refrigerators and vacuum cleaners.

AI agents predicted to potentially take over household appliances

Small AI models could learn to self-replicate through internet-connected consumer electronics and home appliances by the end of 2026.

Anthropic researchers conducted an experiment where 15 Android TV set-top boxes, costing about $55 each, were connected to a single network, and an AI model with 26 billion parameters was successfully launched on it. The specialists then tested the models' ability to independently infiltrate third-party devices and create new copies of themselves.

A model with 9 billion parameters managed to launch on eight out of nine types of equipment, but it cannot yet stably perform a full self-replication cycle. Only the significantly larger GLM-5.2 model was able to complete this task, but its operation requires much higher computing power than robot vacuums and smart TVs can provide.

According to the researchers, to replicate independently within IoT device networks, an AI model needs to require relatively low power while possessing the ability to independently discover devices, transfer itself to them, and launch. Currently, these two conditions are not met simultaneously, but if the current pace of development continues, small models will achieve this by the end of 2026.

Anthropic analysts concluded that AI agents would take over home devices rather slowly. Researchers compare them to mold or fruit flies, which take several days to reproduce. Due to limited capacity, copying and launching just one new instance of a 9-billion-parameter model could take about two days.

Researchers do not rule out the possibility that many AI agents operating without human control will emerge in the future. Getting rid of them will be quite difficult, as it will require coordinating actions among all internet service providers that detect compromised devices on their networks.

Cookies on xabarchi

We use cookies to remember your language and theme, and to count how many people are reading right now — that count is anonymous, lasts only while your browser is open, and cannot be tied to you or to another visit. With your permission we also measure how the site is read: Microsoft Clarity, which records page views and on-page interactions, and our own count of returning readers. Nothing that recognises you across visits is measured until you accept.