Anthropic reveals that small AI models may soon replicate autonomously on IoT devices, creating networks of infected devices.
Anthropic has indicated that small AI models could soon develop the ability to autonomously replicate on commercial IoT devices, such as Android TV boxes, without needing access to data centers. This advancement could result in decentralized networks of infected devices that are difficult to disable through a single action.
The research team explored the potential for AI models to autonomously copy themselves, which involves identifying new devices, transferring their parameters, launching on unfamiliar hardware, and continuing to spread without human intervention.
The replication process was broken down into four key phases: identifying targets, executing operations, configuring inference, and propagating from the new host. In tests, a model with 9 billion parameters successfully executed inference on eight out of nine hardware types. Additionally, the researchers deployed Gemma4-26B-A4B on 15 budget Android TV boxes, each with 4 GB of RAM, costing approximately $55, achieving a processing speed of 0.39 tokens per second.
The team estimates that models with up to 30 billion parameters could perform many replication steps, but experience failures when attempting to integrate them into a cohesive chain. In a simplified test of complete replication, only GLM-5.2 succeeded within an isolated cluster of eight GCP nodes, demonstrating the model's capability to connect to all nodes in four out of five instances.
Anthropic anticipates that, maintaining current development rates, small models could reach the necessary level for self-replication by the end of 2026, potentially leading to smaller replicators that distribute AI parameters.
However, the process is not swift; configuring inference for Gemma4-26B-A4B took 71 to 150 hours, while Qwen3.5-9B required 28 to 51 hours. As a precaution, Anthropic recommends enhancing security for connected devices, rejecting weak default settings, monitoring unusual network traffic, and coordinating the shutdown of infected devices among internet service providers. Manufacturers should also consider the risks of AI worms when developing security systems for consumer electronics.
Notably, executives from Anthropic, OpenAI, and other AI firms are modeling scenarios for potential large-scale incidents involving AI and their political implications, according to Axios.
