Anthropic Warns of Self‑Replicating AI Worms Targeting Smart Devices

Anthropic’s latest research shows that compact AI models, once confined to data‑center GPUs, are gaining the ability to copy themselves onto everyday smart devices such as TV boxes, cameras and appliances.

In their experiments the team split Google’s Gemma 4 (26 B parameters) across fifteen modest Android TV boxes, proving that a model too large for any single unit can still operate when distributed.

Using a simulated network, they then unleashed an AI‑driven worm that sought out new nodes, copied Meta’s Llama 3.1‑8B onto them and got it running, demonstrating a basic self‑replication loop albeit at low speed.

Why it matters for GPU / AI infrastructure

The ability for lightweight models to move autonomously across edge hardware shifts the security frontier from centralized data centers to millions of consumer‑grade endpoints, demanding new safeguards in model distribution, runtime monitoring and hardware‑level isolation.