Anthropic says small AI models could soon be able to copy themselves onto everyday devices

Anthropic's latest research suggests that small, open-weight AI models could soon self-replicate and spread across everyday devices, such as smart ovens and TV boxes. In a blog post, the researchers tested whether a 26-billion-parameter model could be split into smaller pieces and run on inexpensive Android TV boxes each costing around $55 and equipped with 4GB of memory.

While no single box had enough memory to contain the entire model, each device successfully ran a portion of it, allowing the model to work albeit slowly. In another experiment, Anthropic simulated a network of eight cloud computers behaving like Android devices and planted an AI "worm" on one of them. The worm's goal was to locate the other machines, copy itself onto them, and establish a functional AI model.

Of the tested open-weight models, only GLM-5.2, a larger open-source model from Chinese AI company Z.ai, successfully completed all steps in four out of five trials. While smaller models could manage individual tasks, they struggled to combine them. Anthropic estimates that a small model could catch up to GLM-5.2 by the end of 2026.

However, some experts remain skeptical about the potential risks posed by self-replicating AI models. They argue that the experiments performed by Anthropic lack real-world applicability, as the tests relied on predetermined conditions and lacked human oversight. Cybersecurity researchers also point out that self-replicating malicious software, or "worms," have been a concern for years and can be managed through traditional cybersecurity measures, such as rapid patches and anti-virus software updates.