The key facts

  • Anthropic ran a 26-billion-parameter AI model on 15 Android-TV boxes costing about $55 each to test self-replication.
  • A 9-billion-parameter model ran on 8 of 9 device types but cannot yet complete full self-replication; only GLM-5.2 could, needing more power than such devices offer.
  • If current development pace holds, small AI models may master autonomous replication on IoT networks by the end of 2026, spreading slowly — roughly one copy per two days.

Small AI models may master self-replication via internet-connected consumer electronics and home appliances by the end of 2026. This is the conclusion reached by Anthropic. conclusion .

Anthropic researchers conducted the following experiment: 15 Android-TV boxes, each costing about $55, were united into 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 foreign devices and create new copies of themselves.

A 9-billion-parameter model managed to run on eight of nine types of equipment, but it cannot yet reliably perform the full cycle of self-replication. Only the considerably larger GLM-5.2 model was able to accomplish this task, though running it requires far more computing power than robot vacuum cleaners and smart TVs can provide.

According to the researchers, for an AI model to replicate autonomously across IoT device networks, it must require relatively little power, yet at the same time be capable of independently finding devices, copying itself onto them, and launching itself. At present 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 will take over home devices rather slowly. The researchers compare their spread to mold or fruit flies, whose reproduction takes several days. Due to limited power, copying and launching even a single new copy of the 9-billion-parameter model could take about two days.

The researchers do not rule out that in the future many AI agents operating without human oversight may emerge. Getting rid of them will be quite difficult, since it would require all internet providers in whose networks infected devices are detected to coordinate their actions with one another.