Anthropic Warns Small AI Models May Soon Be Capable of Self-Replication on Smart Devices

Anthropic has warned that small artificial intelligence models could soon spread across internet-connected devices without human assistance, following experiments examining AI replication and security risks on common smart devices.

Anthropic Research Finds AI Models Near Self-Replication Capability

According to Anthropic’s research, smaller AI models could gain the ability to copy themselves across connected devices within months. During the experiments, researchers tested whether these systems could identify vulnerable hardware, install software, and operate without human intervention.

Models with fewer than 30 billion parameters completed several steps needed for replication during the experiments. But they were not able to complete the whole process, as some of the tasks involved a higher level of decision-making.

GLM-5.2 and stronger AI models successfully completed the full replication process under simplified testing conditions. Based on current development trends, Anthropic expects smaller models could reach similar abilities by the end of 2026.

The results follow the company’s increased research focus on artificial intelligence. Recently, Anthropic committed $150 million to the US Genesis Mission, supporting wider use of artificial intelligence in scientific research.

AI Models Run Across 15 Android TV Devices

During the experiments, researchers successfully operated an AI model across 15 low-cost Android TV boxes without using traditional data centers. The cost of each device was around $55 and the combined hardware supported the Gemma4-26B-A4B model.

Another test showed that a nine-billion-parameter model could run on eight of nine different hardware types. The demonstration was a showcase of the potential for smaller AI systems to function on more common devices, even with limited computing power.

Smart televisions, refrigerators, ovens and robotic lawnmowers were cited as potential targets because many have little to no security built-in. These devices could provide computing resources for AI systems attempting to spread through connected networks.

However, the recent increase in the use of autonomous software has made its way into financial markets, such as AI agents trading cryptocurrencies. These types of trading systems are not replicated in an unregulated fashion as in the scenarios examined by Anthropic.

Anthropic Warns AI Worms Could Be Difficult to Stop

Anthropic said that self-replicating AI systems don’t need a central server to replicate and therefore are more difficult to turn off. In addition to traditional malware, these systems might change their tactics on seeing an unknown device.

Initial releases would probably take days to double the number of devices in a mutually reachable set of hardware. However, they may be able to work on multiple devices without human intervention.

Anthropic recommended enhanced security of devices and more monitoring of connected networks. The company also noted that none of the tested small model has been capable of self-replication on real-world smart devices without restriction.

For more investment opportunities, retail investors can explore pre-IPO tokens offering exposure to major private tech companies.