Anthropic: AI models may soon copy themselves across cheap smart devices Anthropic

Anthropic's Frontier Red Team says AI models may soon be able to copy themselves across cheap smart devices, like TV boxes and cameras, with no datacenter and no GPUs. A network like that would have no centre, which makes it hard to shut down. Today the risk is low: replication is slow and shows up on the network. But small models are getting better fast.

The experiment

• Cheap hardware is enough to run a model. The authors bought 15 off-brand Android TV boxes at about $55 each, linked them, and ran Gemma4-26B-A4B across them, at 0.39 tokens per second.
• 9 kinds of real hardware were tested; 8 of the 9 ran a 9B model.
• Can an agent spread on its own? On a chain of 8 virtual Android devices, GLM-5.2 infected all 8 nodes in 4 of 5 runs, and 7 of 8 in the fifth. What it copied and ran on each node was the smaller Llama3.1-8B, not itself: GLM-5.2 is too big for a TV box.
• It is slow. One replication took 71 to 150 hours with Gemma4-26B-A4B, 28 to 51 hours with Qwen3.5-9B.

Why it matters

GLM-5.2 and stronger models can already do the whole job. A model small enough to live on a set-top box cannot yet. Anthropic estimates that if small models keep improving at their historical rate, they will match this level by the end of this year. At that point the model could copy itself, not a smaller helper.

Anthropic compares such a replicator to "a slime mold or a swarm of fruit flies, which replicate over days": slow, quiet, and with no single place to switch off.

What can stop it

• Watching networks for unusual traffic between devices.
• Internet providers quarantining infected devices: this was 92% effective against the Mirai botnet.
• Rules that make smart devices ship with secure defaults.

Sources

• Anthropic: AI models may be capable of self-replicating without access to a datacenter (Frontier Red Team, October 12, 2026)