Anthropic published a report titled "AI models may be capable of self-replicating without access to a datacenter" on October 12, 2026.

In tests conducted within a firewalled VPC of 8 GCP nodes running redroid, an Android 15 emulator, the GLM-5.2 model spread to 8 of 8 nodes in 4 out of 5 trials and to 7 of 8 nodes in 1 trial. Researchers decomposed this autonomous replication process into four stages: finding targets, gaining execution, establishing inference, and continuing propagation from the new host.

The team also ran the Gemma4-26B-A4B model distributed over 15 off-brand Android TV boxes costing approximately $55 each. This distributed setup achieved an inference speed of 0.39 tokens per second across the devices. Researchers estimated that unoptimized inference setup for this model takes between 71 and 150 hours.

Evaluations of other open-weight models showed that a 9B parameter model successfully ran inference on 8 of the 9 hardware types tested, which included Jetson Nanos, Mac PCs, and RTX 4090s serving Llama3.1-8B. For the Qwen3.5-9B model, unoptimized inference setup was estimated to take between 28 and 51 hours. In experiments, the inference setup success rate for Qwen3.5-9B rose from 45% in the first generation to between 61% and 84% in later generations through inherited notes, though giving the model access to Opus 4.8 as an advisor did not measurably help success. To mitigate failure modes in these long-horizon runs, researchers used a harness called Memento, built on top of the open-source Pi coding agent.

Why It Matters

The report notes the vulnerability and scale of consumer devices by noting that the Kimwolf botnet infected over 2 million devices, primarily Android TV boxes. Researchers calculated that a botnet of 2 million TVs serving AI models would yield 80,000 tokens per second across the network. The study also pointed to mitigation strategies, noting that quarantining infected subscribers cleared 92% of Mirai botnet infections within 14 days, compared to 74% with no notification.