Anthropic reported experiments suggesting that small AI models may be months away from being capable of self-replication on commercial IoT devices like smart ovens and off-brand Android TV boxes.[1] According to Anthropic, in their evaluations, a 9-billion-parameter model successfully ran inference on 8 of the 9 hardware types tested.[1] Models of 30 billion parameters and under can already complete many individual steps for replication, though they make mistakes when chaining them end to end, the company said.[1] In our simplified evaluations, GLM-5.2 and stronger models are capable of end to end replication, and if small models improve at their historical rate, they’ll match this capability level by the end of this year, according to Anthropic.[1]
Unlike non-AI malware, it could adapt to unfamiliar devices on its own, according to Anthropic.[1] Anthropic said that weight-based replicators are autonomous worms that run their models directly on compromised hardware, and that the weights would become part of the malware payload.[1] Anthropic ran Gemma4-26B-A4B distributed over 15 inexpensive off-brand Android TV boxes (~$55 each), achieving 0.39 tokens per second, according to Anthropic.[1]
