Anthropic warns self-replicating AI could reach IoT devices within months

Anthropic warns that small AI models could soon replicate across internet-connected devices without relying on a data center. The company says these systems could move between poorly defended Internet of Things (IoT) hardware and run inference across compromised devices.

To test the idea, Anthropic researchers ran Gemma4-26B-A4B, described as a 25 billion-parameter model, across 15 inexpensive Android TV boxes that cost about $55 each. Connected as a group, the boxes ran distributed inference, demonstrating a proof of concept.

The researchers also found that Qwen3.5-9B adapted to 8 of the 9 hardware types they tested, including Jetson Nanos, Mac PCs and RTX 4090s. Anthropic says smaller models may adapt more readily to unfamiliar hardware, while larger models may replicate faster.

The company says self-replicating AI is not an immediate threat, though its researchers theorize it could become a problem by the end of the year. Potential warning signs include large model payloads sent to newly infected devices and unusual traffic between machines. Anthropic identifies network quarantine, approved-site whitelists, stronger IoT security and longer-lasting automatic updates as possible defenses.