Can AI spread copies of itself? Considering preparations based on Anthropic's 'self-replication' evaluation

The research also ran a chain‑replication test on eight Android 15 emulators in a Google Cloud isolated network. In four of five trials, a GLM‑5.2 model reached all eight devices, showing that models can combine the tasks needed for replication into a single workflow. The team notes that the evaluation used simplified conditions—no authentication, accessible ADB ports—and that real‑world attacks would be more difficult.

Anthropic recommends basic security hygiene: change default passwords, apply updates, disable unused features, and monitor unusual inter‑device traffic. The findings suggest that while autonomous AI worms are not yet confirmed, small models are gaining the ability to self‑replicate in controlled environments.