Anthropic Tests Self-Replicating Open AI Models on Cheap IoT Hardware

On October 12, researchers William Zhou and Daniel Freeman from the Anthropic Frontier Red Team published an experiment evaluating whether compact open-weights models could autonomously replicate across low-cost connected hardware without relying on data centers.

Testing in a simplified network environment, a 9-billion-parameter model managed to configure inference across eight of nine hardware configurations. In trials using GLM-5.2 from Zhipu, the system spread across all eight test nodes in four out of five runs, and reached seven out of eight in the fifth. In another trial, Google's Gemma4-26B-A4B ran across 15 unbranded $55 Android streaming boxes, generating text at a sluggish 0.39 tokens per second.

While that speed is impractical for interactive chat, autonomous botnets do not have human deadlines. The researchers noted that bootstrapping inference on the TV boxes took 71 to 150 hours for Gemma and 28 to 51 hours for Qwen3.5-9B. Interestingly, later generations of Qwen3.5-9B improved their success rate: by passing configuration notes to subsequent copies, setup reliability jumped from 45% in the initial run to between 61% and 84% in later iterations.

The test exploited an unauthenticated, open ADB port, which the authors explicitly highlighted as an artificial, best-case scenario for propagation. Extrapolating to the scale of real-world outbreaks like the two-million-device Kimwolf botnet, grouping devices into clusters of 15 running Gemma would theoretically yield around 133,000 active model instances and an aggregate throughput of roughly 80,000 tokens per second—roughly matching 40 Nvidia H100 GPUs, but scattered across domestic living rooms.

The researchers cautioned that only GLM-5.2 operated near the performance threshold needed for full end-to-end self-replication, projecting that similar small models could bridge that gap by late 2026 if current capability growth holds.

Source: habr

Primary source: Anthropic