Anthropic Experiments: Risk of Autonomous AI Model Replication

Anthropic researchers conducted a series of experiments to test the ability of language models to self-replicate without using centralized data centers. During testing, they evaluated whether models could copy their weights to third-party devices and run inference there.

Autonomous Deployment on Various Hardware

In the first experiment, the Qwen3.5-9B model independently configured an inference server launch. It successfully deployed on 8 out of 9 tested device types, including Jetson Nano, Mac, RTX 4090 GPUs, and AMD MI300X.

Gathering Computing Resources from Low-Cost Devices

The Gemma4-26B-A4B model was split into parts and run on 15 Chinese Android TV boxes costing $55 each. The combined performance reached 0.39 tokens per second, demonstrating the feasibility of serving large models using IoT hardware.

Infection Cycle via Open Ports

The GLM-5.2 model used open ADB ports to scan for nodes, copy payloads, and deploy Llama3.1-8B. It traversed all 8 network nodes in 4 out of 5 attempts, demonstrating autonomous network movement capabilities.

Limitations of Current Technology

Creating a full-fledged AI virus remains difficult due to large model weights and lengthy deployment times. GLM-5.2 weighs over a terabyte and requires 28 to 150 hours to install on a single box, whereas traditional worms infect machines in minutes.

Threat Projections

Anthropic predicts that models with 30 billion parameters will reach the necessary level of autonomy by the end of 2026. The experiments highlight the importance of checking port security on home devices like Android TV boxes.