Anthropic Warns of the Self-Replication Risk of Small AI Models

The artificial intelligence company Anthropic has raised concerns about the possibility of small AI models self-replicating through Internet of Things (IoT) devices. According to a report by The Verge, Anthropic has conducted experiments showing that these models can spread to household devices like refrigerators, vacuum cleaners, and security cameras, creating a new cybersecurity risk.

Experts have emphasized that while computer worms and the fear of self-replicating AI are not new, recent advancements suggest these warnings should be taken seriously. Small AI models, which are continuously improving and can operate on common devices, present a new challenge as they can spread across networks of less powerful devices.

In one experiment, Anthropic researchers managed to distribute the Google Gemma 4 model, with 26 billion parameters, across 15 Android TV boxes, demonstrating that an AI model can operate on household devices. However, only the most sophisticated model tested, GLM-5.2 from China’s Z.ai, was able to spread across the entire network.

Jeffrey Ladish, executive director of Palisade Research, an AI security research firm, highlighted that these smaller models are capable of conducting cyberattacks autonomously. He predicts that large models could operate on large networks of consumer chips, including those found in MacBooks.

Nicolas Papernot, a professor at the University of Toronto, warned that AI-driven computer worms are a tangible threat and require immediate action, especially to secure critical infrastructure like hospitals and power grids. He stressed that without a massive effort to secure devices, AI worms will have a conducive environment to spread.