Anthropic Warns of Future AI Self-Replicating Worms

Anthropic’s research explored the potential for AI models to self-replicate across commercial IoT devices, finding that while such attacks are not yet practical, the technology is advancing rapidly. The study demonstrated that a network of low-cost Android devices could host an AI model, and some models could autonomously spread across a network. The tests, conducted under highly favorable conditions, showed that GLM-5.2, a more capable model, successfully completed the end-to-end replication process. Researchers concluded that smaller models may soon achieve similar capabilities, posing significant challenges for defenders.