Anthropic researchers William Zhou and Daniel Freeman tested autonomous replication and adaptation (ARA) in a sandboxed eight-node GCP network running Android 15 redroid emulators: from a single seeded node, GLM-5.2 independently scanned unauthorized ADB ports, spread to neighbours, moved model weights and an execution framework, deployed a Llama3.1-8B inference service and spawned a next-generation worm, taking all eight devices in four of five runs and seven of eight in the fifth. Another experiment spread a 26B model (Gemma4-26B-A4B) across 15 second-hand $55 Android TV boxes with 4GB RAM via Petals-style decentralized inference at 0.39 tokens per second, and the write-up estimates that 2 million Kimwolf-botnet-style devices could host about 133,000 AI copies totalling roughly 80,000 tokens per second. A 'Memento' framework in which a Qwen3.5-9B agent inherits debugging notes from failed predecessors raised autonomous deployment success from 45% in the first generation to 61-84% in later ones, and the researchers project sub-30B models could reach GLM-5.2-level autonomy by the end of the year.

Anthropic·GLM-5.2·Gemma4-26B-A4B·+5