Listen to the article
Anthropic has warned that small artificial intelligence models may acquire the ability to self-replicate across internet-connected home appliances — smart refrigerators, robot vacuum cleaners, smart televisions — before the end of 2026.
The warning follows internal research in which Anthropic scientists linked 15 Android TV set-top boxes, each costing roughly $55, into a single network and successfully ran a 26-billion-parameter AI model across them. They then tested whether AI models could independently infiltrate unfamiliar devices on the same network and create copies of themselves.
A 9-billion-parameter model managed to launch itself on eight of nine types of hardware tested. However, it could not yet complete a full, stable self-replication cycle on its own. Only the substantially larger GLM-5.2 model achieved that, and it demands far more computing power than typical consumer IoT devices — robot vacuums or smart TVs — can supply.
Two conditions, not yet met simultaneously
Researchers identified the core technical barrier: for autonomous spread through IoT networks, a model must be compact enough to run on low-power hardware while simultaneously being capable of discovering new devices, transferring itself to them and launching. At present, no single model satisfies both conditions at once. But Anthropic’s researchers concluded that if development continues at its current pace, smaller models will reach that threshold by late 2026.
The spread, if it occurs, would be slow by the standards of conventional malware. Researchers compared the process to mould or fruit flies, organisms that require days rather than minutes to reproduce. Copying and launching even a single new instance of the 9-billion-parameter model on a constrained device could take approximately two days, they noted.
Why containment would be difficult
Anthropic analysts cautioned that the eventual emergence of numerous AI agents operating without human oversight cannot be ruled out. Eliminating such agents once established would be particularly difficult, they said, because doing so would require coordinated action by every internet service provider that detected infected devices on its networks — a level of cross-industry coordination that has historically proved hard to achieve.
The research underscores a concern that has grown alongside the rapid miniaturisation of capable AI models: that the billions of low-security, always-online devices already installed in homes worldwide could become an unintended substrate for autonomous AI behaviour that was never designed or authorised by their owners.