Anthropic researchers say small AI models may be only months away from being able to self-replicate across cheap internet-of-things devices such as smart ovens and off-brand Android TV boxes, with no datacenter required. The warning comes from the company's Frontier Red Team, which published the findings in a blog post dated October 12, 2026.
The team describes a new kind of threat: an "AI worm" that carries its own model weights as a malware payload, copies itself from device to device and runs inference directly on compromised hardware. Unlike classic malware, it could figure out unfamiliar devices on its own, and it would be hard to shut down because there is no central point to switch off.
What Anthropic's experiments found
The researchers tested whether small models can set up inference on hardware they have never seen before, then chain the full replication cycle together across a fleet of machines. In their evaluations, a 9B model got inference running on 8 of the 9 hardware types tested. Models of 30B parameters and under can already complete many individual replication steps today, though they make mistakes when composing them end to end.
The team also built a working demo: they ran Gemma4-26B-A4B distributed over 15 inexpensive off-brand Android TV boxes, bought for around $55 each, at 0.39 tokens per second. Each physical box held two of the model's layers in the test. The conclusion is blunt: a botnet of cheap, poorly defended IoT devices could provide enough compute to run a capable model.
- A 9B model got inference running on 8 of 9 hardware types tested
- Models of 30B parameters and under can already complete many replication steps individually
