Anthropic researchers reported on Monday that small AI models might be able to self-replicate on insecure smart home devices.

The research studied behaviour of small models including Z.ai's GLM-5.2, Alibaba's Qwen and Google DeepMind's Gemma4-26B-A4B. Named devices include smart ovens, refrigerators and lawn mowers.

Work filed 12 October on arXiv identified S-Space, a linear representation subspace in multimodal language models. The listed research team includes Zhiyuan Shi, Xiang Li and nine additional authors.