Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

AI risk watchWarnings from authoritative sources

Anthropic: small open models could copy themselves across cheap smart-TV boxes by the end of 2026

AI labLoss of control

Anthropic tested whether small open-weight models can install and spread themselves across off-brand Android TV boxes. A 26-billion-parameter Gemma ran spread over 15 boxes; Qwen3.5-9B set up inference for another model on eight of nine hardware types; and in an emulated network of eight devices with unauthenticated ADB, GLM-5.2 alone, of 11 open models tested, completed the full replication cycle to every node, in four of five trials. Keeping what it learned raised Qwen's success from 45 to 84 percent. No autonomous replicator exists yet and spread would double in days, but Anthropic expects models of 30 billion parameters or fewer to reach GLM-5.2's level by the end of 2026, and asks ISPs and device makers to prepare: detect large model transfers, watch device-to-device traffic, run honeypots, quarantine infected subscribers, and ship devices without default passwords.

Why it matters

A botnet that carries its own model has no command server to seize. The window to build the network-level detections Anthropic describes is the months before the models get there.

What you can do

  • Inventory the IoT and embedded devices on your networks, close ADB and other debug services, and change default passwords.
  • Alert on large transfers between devices that never talked to each other before.
  • Treat a compromised device that starts talking to its neighbors as an incident, not noise.

Sources

See it with the other risks in the Learning Center