Anthropic researchers test whether AI models can autonomously replicate across cheap devices

Anthropic researchers William Zhou and Daniel Freeman tested whether GLM-5.2 could autonomously spread an 8B-parameter model across eight isolated devices. GLM-5.2 fully compromised all eight devices in four of five runs and seven in the fifth.

The author also describes a distributed test in which a 26B model ran across 15 low-cost Android TV boxes at 0.39 tokens per second, and the researchers estimate that sub-30B models may reach end-to-end replication by the end of 2026.