No. Anthropic’s October 12, 2026 report described distributed inference across 15 physical Android TV boxes and a separate propagation evaluation on Android emulators—not an AI worm infecting a household device. The two tests had different goals and environments, and Anthropic’s report gives the details behind both.
The two tests did different jobs
| Test | Model and task | Environment | Reported result |
| Distributed inference | Gemma4-26B-A4B split across devices | 15 off-brand Android TV boxes, each with 4 GB RAM; researchers paid about $55 per box | 0.39 tokens per second |
| Propagation evaluation | GLM-5.2 coordinated the chain; successful nodes served Llama3.1-8B | Eight Android 15 emulators in a firewalled Google Cloud Platform virtual private cloud, with an unauthenticated ADB port | All eight nodes reached in four of five trials; seven of eight in the fifth |
The physical demonstration showed how computation could be distributed across a cluster of boxes. It did not put the full model on one ordinary smart-home device: under Anthropic’s 4-bit quantization assumption, inference required at least four machines for a 9B model and ten for a 26B model. In a separate hardware evaluation, Qwen3.5-9B established inference on eight of nine tested hardware types. Anthropic estimated inference-setup times of 28–51 hours for Qwen3.5-9B and 71–150 hours for Gemma4-26B-A4B in its described workload.
Why the propagation result has a narrow scope
The propagation evaluation used Android 15 emulators, not commercial appliances. Its ADB port—the Android Debug Bridge interface used to connect to the devices—had no authentication. Anthropic said that made the cyber-exploit step trivial and the result an upper bound on success against real-world security measures.
The model roles matter, too. GLM-5.2 coordinated the propagation, while the reached nodes served Llama3.1-8B; the test did not show GLM-5.2 running on every target. Anthropic also reported that models with 30 billion parameters or fewer could complete many individual steps but made mistakes when combining them into a full chain.
What defenders could monitor
Anthropic proposed watching for large model-payload transfers, unusually sustained traffic between devices, network scanning and unexpected outbound connections. It also discussed honeypots, ISP quarantine and stronger default security for connected devices. These are proposed defensive measures, not results from an AI-worm response in the field.
As a separate precedent, Anthropic cited Mirai remediation in which ISP quarantine and customer notifications cleared 92% of infections within 14 days, compared with 74% without notification.
The forecast was conditional
Anthropic estimated that smaller models could reach the evaluated capability level by the end of 2026 if they continued improving at their historical rate. The company described that as a rough forecast and the capability index behind it as a proxy for replication ability.
