Anthropic Frontier Red Team Runs 26B-Parameter Model at 0.39 Tokens/Second Across Eight $55 Android TV Boxes

The team tested whether models could perform inference on unfamiliar IoT hardware rather than relying on prior exploitation research. A 9-billion-parameter model succeeded on eight of nine device types. Only GLM-5.2 completed the full worm cycle, reaching all eight emulated Android devices in four of five runs on a closed network with trivial initial access. Researchers deliberately used default credentials to isolate the inference and replication variables.

Scaling data shows a two-million-device botnet like Mirai would host roughly 133,000 model copies on equivalent hardware, delivering at most 40 H100 GPU equivalents with each replication cycle exceeding one day. Models under 30 billion parameters already handle isolated steps but fail when chaining actions; historical improvement rates project reliable chaining by late 2026.

Open-weight diffusion combined with persistent default-password exposure on millions of IoT devices creates a distributed inference substrate that lacks any central kill switch. ISP monitoring of sustained inter-device model transfers and large binary movements offers the clearest near-term detection signal, while manufacturers shifting away from factory credentials would shrink the available target pool faster than capability growth.

Next observable milestones include public replication of the eight-device closed-network result on real heterogeneous hardware and measurable error-rate reductions in step-chaining for 20-30B models within 18 months.