Anthropic: small AI models could self-replicate on IoT devices without GPUs

Anthropic: small AI models could self-replicate on IoT devices without GPUs

Anthropic claims that small AI models may soon be able to self-replicate on IoT devices without access to data centers or GPUs.

The scenario describes AI learning to infect household electronics such as cheap smart TV boxes, cameras, and other smart devices, remain on them, and spread to further devices without human involvement. Such a network is nearly impossible to shut down because it has no central point.

Physically, even simple hardware can run a small model. In a demo, the authors bought 15 Android TV boxes for about $55 each, connected them, and successfully ran Gemma4-26B-A4B on them.

The main question is different: can an agent, placed on unfamiliar hardware, figure out how to run inference? The researchers tested this on 9 types of real hardware and on a chain of 8 virtual devices.

As a result, GLM-5.2 infected all 8 nodes in 4 out of 5 runs. So far it did not launch itself, because it is too large, but ran Llama 8B. Still, once a sufficiently small model with similar capabilities appears and fits on a TV box, it could already self-replicate.

For now, no such models exist, and full inference setup can take up to 150 hours, so the risk is limited. However, Anthropic compares the virus to slime mold or fruit flies: it can reproduce over days, very quietly, and is extremely hard to eliminate because there is no center.

Anthropic research