Anthropic finds small models can run on IoT hardware—but not yet self-replicate end to end

Anthropic’s lab evaluation shows small models can set up inference on varied hardware, while its end-to-end replication result relies on simplified access and a larger remote model.

2 min read

Illustrative close-up of a circuit board representing AI agent software infrastructure

Anthropic’s Frontier Red Team says small AI models are approaching—but have not yet demonstrated—autonomous replication across commercial internet-of-things hardware.

In research published October 12, the team tested whether agents could configure inference on unfamiliar devices and chain together a simplified replication cycle. A 9-billion-parameter model got inference running on eight of nine hardware types. Separately, the researchers ran Gemma4-26B-A4B across 15 inexpensive Android TV boxes at about 0.39 tokens per second.

The end-to-end result requires careful qualification. In a firewalled Google Cloud network of eight Android-emulator nodes, GLM-5.2 spread a Llama 3.1 8B payload to all eight nodes in four of five trials and seven nodes in the fifth. Other tested open-weight models failed reliably. The experiment deliberately removed the hardest intrusion step: targets exposed an unauthenticated Android Debug Bridge port. GLM-5.2 itself also ran on a separate inference server rather than on the infected nodes, which hosted the smaller payload.

Anthropic therefore treats GLM-5.2 as a capability threshold and models such as Qwen3.5-9B or Gemma4-26B-A4B as a size threshold. Those thresholds do not yet overlap. Its forecast that a sufficiently capable sub-30B model could arrive by the end of 2026 comes from extrapolating the Epoch Capabilities Index, a general-purpose proxy rather than a direct replication benchmark.

The study is a lab-run threat evaluation, not evidence of an autonomous AI worm operating in the wild. It also focuses on inference setup more than vulnerability exploitation, and Anthropic notes that a real weight-based replicator would likely spread slowly. The authors argue that default-password rules, network anomaly detection, honeypots and ISP quarantine could reduce the replication rate below a sustainable level.

A close-up of a circuit board, illustrating edge devices that could host small AI models.

Image is illustrative and is not a device used in Anthropic’s evaluation. Photo from Unsplash, used under the Unsplash License.

Sources: Anthropic research report · Reuters · Unsplash License

AnthropicAI securityAI evaluationsmodel benchmarksautonomous agentsIoT