Anthropic Flags Risk of Self-Replicating AI Spreading Across Smart Devices

Anthropic has published research describing how compact AI models might one day copy themselves across everyday connected devices, forming distributed networks capable of running AI workloads without data center infrastructure. The findings were covered by BleepingComputer and Tom's Hardware.

The company's Frontier Red Team detailed the concept in a report dated October 12, titled "AI models may be capable of self-replicating without access to a datacenter." The paper examines what it calls a weight-based replicator: a type of malware that, instead of spreading conventional executable code, would distribute AI model weights together with the software needed to run inference on compromised hardware.

According to Tom's Hardware, the researchers demonstrated two building blocks of such a system. They showed that AI agents can set up inference across varied hardware platforms and coordinate replication over a network. In testing, the Qwen3.5-9B model was able to establish an inference server on eight of nine hardware types examined, including NVIDIA Jetson Nano devices, Apple Silicon Macs, Xeon servers, a Google TPU, AMD's MI300X, and a PC with an RTX 4090.