Anthropic said in October 2026 that AI self-replication leaves detectable network signatures such as multi-gigabyte payload transfers and device-to-device inference traffic

Overview Debates Related claims Topics Sources

From the postAI models may be capable of self-replicating without access to a datacenter

AI models may be capable of self-replicating without access to a datacenter

Anthropic researchers show small open-weight AI models can set up inference on unfamiliar IoT hardware and, in simplified evals, GLM-5.2 can replicate end to end. They project small-model worms by end of 2026 and suggest defenses.