AI Could Soon Replicate Across Millions of Smart Devices, Anthropic Warns

Anthropic warns that smaller AI models could soon gain the ability to replicate across smart devices, potentially creating decentralized AI networks without relying on traditional data centers.

According to Anthropic’s research, AI models are already showing the ability to identify compatible hardware, connect devices into networks, and configure software needed to run additional AI models.

Researchers estimate that models with 30 billion parameters or fewer could develop self-replication capabilities by the end of 2026, although current experiments have not demonstrated fully autonomous replication on real IoT hardware.

AI models can already spread workloads across smart devices

Anthropic tested whether AI models could use inexpensive consumer hardware, including Android TV boxes, to create distributed computing networks.

The experiments produced several notable results:

• Gemma4-26B-A4B: Researchers successfully distributed Google’s AI model across 15 inexpensive Android TV boxes.
• Qwen3.5-9B: The model installed and configured software needed to run another AI model on eight of nine hardware types tested.
• GLM-5.2: The model reached all eight devices in an emulated network in four out of five trials.

Qwen3.5-9B also improved its success rate from 45% to 84% by reusing information from previous attempts.

However, researchers conducted these experiments under simplified security conditions. Some devices allowed unauthenticated Android Debug Bridge (ADB) access, making them easier to configure or compromise than properly secured consumer devices.

Self-replicating AI could emerge before 2027

Anthropic estimates that smaller AI models could reach the necessary capabilities by late 2026 if they continue improving at historical rates.

However, replication remains relatively slow on current hardware.

Researchers estimate that setting up Qwen3.5-9B across tested devices would take approximately 28 to 51 hours, while Gemma4-26B-A4B would require 71 to 150 hours.

These delays could give security teams enough time to detect unusual activity and isolate affected devices.

The bigger concern is that future models could complete these tasks faster while distributing themselves across hundreds or thousands of vulnerable devices.

Unlike traditional AI services, such networks would not depend on a single central server, potentially making them harder to disable.

Anthropic recommends stronger IoT security

Anthropic warns that vulnerable smart devices, including Android TV boxes and connected household appliances, could eventually provide computing resources for unauthorized AI networks.

To reduce the risk, researchers recommend monitoring unusual traffic between devices, detecting large AI model transfers, deploying honeypots, and quarantining compromised hardware.

Improving default IoT security could also prevent AI systems from exploiting poorly configured devices.

For now, Anthropic emphasizes that the threat remains a potential future risk rather than an immediate danger. Its experiments did not demonstrate reliable, autonomous end-to-end replication across real-world devices.

In other security news, OpenAI used its own AI to trace a hidden recording device, while Anthropic cut AI internet access after models targeted government websites. Meanwhile, Google Chrome 155 fixes 37 security vulnerabilities, including seven critical flaws.