NewsResearch
Anthropic: small open models are months from copying themselves across cheap smart-TV boxes, and defenders should prepare now
Anthropic bought off-brand Android TV boxes and tested whether small open-weight models could install and spread themselves across them, BleepingComputer and SecurityWeek reported on October 13. Gemma4-26B-A4B ran spread over 15 boxes; Qwen3.5-9B, acting as an agent, installed and configured inference for Llama3.1-8B on eight of nine hardware types, taking 28 to 51 hours per setup. In an emulated network of eight Android devices with unauthenticated ADB, a deliberately simplified intrusion step, only GLM-5.2 among 11 open models from Alibaba, DeepSeek, Google, Meta, Moonshot, OpenAI and Zhipu completed the whole replication cycle to all eight nodes, in four of five trials; letting Qwen keep what it learned raised its success from 45 to 84 percent, while advice from Opus 4.8 made no measurable difference. Anthropic says no fully autonomous replicator exists, that spread would double in days rather than hours, but that models of 30 billion parameters or fewer could reach GLM-5.2's level by the end of 2026, and it proposes that ISPs detect large model transfers, watch device-to-device traffic, honeypots, quarantining infected subscribers and better IoT defaults.
Why it matters
A botnet that carries its own model needs no command server to take down. The defenses Anthropic proposes are network ones: notice a large model file moving between devices that never talked before, and ship devices without default passwords.