Anthropic Warns Self-Replicating AI Worms Are Close
Anthropic’s Frontier Red Team warns that autonomous AI worms able to spread across internet-of-things devices without any data center are becoming feasible. In one experiment the team ran distributed inference on Gemma4-26B-A4B across 15 budget Android TV boxes, a cluster that produced 0.39 tokens per second, and showed that open models can already pick packages and build execution environments on heterogeneous hardware without root access. In an isolated environment GLM-5.2 completed a full replication cycle: scanning the network, compromising a target host, deploying a runtime, and launching a copy of itself. Anthropic estimates that distributed botnets built on models of up to 30 billion parameters will be technically possible by the end of this year.
Related: Anthropic’s Mythos Finds Mathematical Weaknesses in Cryptographic Algorithms, Anthropic Finds Chinese GLM-5.3 Nears Closed Models on Cybersecurity Tasks, Manifold Researcher Discovers Cryptomining Botnet on ClawHub