The "AI Worm" Paper

The Frontier Red Team's October 12 paper ran the experiments: Gemma4-26B-A4B fragmented across 15 cheap Android TV boxes (~$55, 4GB RAM) to run inference at 0.39 tokens/sec; Z.ai's GLM-5.2 completed a full replication sequence on 8 emulated Android nodes in 4 of 5 trials — but only through an unauthenticated ADB port, not over the public internet. A 9B-parameter model established inference on 8 of 9 hardware types. Replication is slow — 28–51 hours per node — a distributed "digital slime mould" doubling in days.

Key facts

Authors: Anthropic Frontier Red Team researchers William Zhou and Daniel Freeman; paper published October 12, 2026; press wave October 14–15.
Claim: small (≤30B) open-weight models are "months away" from self-replicating across commercial IoT devices (smart TVs, cheap Android TV boxes, appliances, cameras); threshold could arrive by end of 2026 at historical improvement rates.
The worm would carry model weights as the payload, adapt to unfamiliar hardware on its own, and have "no centralized point to shut off."
What was NOT shown: a worm spreading over the public internet — the GLM-5.2 replication used an unauthenticated ADB port on emulated nodes.