Anthropic Tests Show AI Worm Risk, Not Real-World Spread

Anthropic researchers say compact artificial intelligence models could become capable of copying themselves across poorly secured connected devices, but their tests did not demonstrate a real-world self-spreading AI worm. The company’s Frontier Red Team said that, if recent capability gains continue, smaller models could reach the level tested by the end of 2026. In research published October 12, the team examined whether models could find devices, install themselves, run on unfamiliar hardware and continue spreading without human direction. Researchers ran a model split across 15 inexpensive Android TV boxes, showing that the hardware could host AI without a data center. The system operated at about 0.39 tokens per second, a slow but functional rate. The researchers also tested a larger open model, GLM-5.2, in a simplified network of eight emulated Android devices. It reached all eight in four of five attempts. The test made the initial break-in unusually easy by leaving an open port that required no authentication, and the larger model ran from a server rather than on the TV boxes.