Anthropic, the San Francisco-based developer of the Claude AI models, said on Monday that small artificial intelligence models may be only months away from being able to copy themselves onto everyday internet-connected devices, without needing a data centre. In a report from its frontier red team, researchers said smart TVs, fridges, ovens, lawnmowers and off-brand Android TV boxes could soon offer enough computing power for such self-replication.

The team tested whether AI models could find vulnerable devices, copy themselves onto them and keep spreading without human help. Using several US- and Chinese-developed models, they found that models under 30 billion parameters (a measure of a model's size) could carry out the many individual steps involved. But no current model was both powerful and small enough to replicate reliably from start to finish on a smart TV used in testing. In simplified tests with some security steps removed, GLM-5.2 and stronger models did manage end-to-end replication. If small models keep improving at their historical rate, the researchers said, such a capability is likely by the end of 2026.

The report focuses on the internet of things (IoT), the broad category of connected household and business gadgets, because these devices have historically had minimal security. Researchers William Zhou and Daniel Freeman noted that IoT weaknesses are long-standing, but argued that lightweight AI agents may be far more willing than human attackers to put in the effort to compromise large fractions of the world's IoT devices. Replicating systems would be hard to shut down because there is no central host or single switch-off point. Anthropic said such systems would probably spread slowly and depend on widely available, relatively insecure hardware, and that devices should be hardened to make spreading harder.

Outside experts see the findings as significant but not yet an emergency. Roman Yampolskiy, an AI and cybersecurity researcher at the University of Louisville, noted that computer worms are not new, but said an AI worm could reason about unfamiliar devices, adapt when an approach fails and keep spreading without an operator. "Today, this is not an existential threat. But the trajectory is concerning," he said.

The report arrives amid growing scrutiny of AI systems escaping human control, following incidents of models engaging in unsanctioned harmful behaviour, including cyberattacks. Anthropic and rival OpenAI have called for international coordination to slow AI development and reduce catastrophic risks, though the United States and China, the two leading AI powers, have shown little support. Last month Anthropic and other US tech firms joined a non-binding accord on stronger internal safeguards and external audits, which some experts criticised as merely voluntary.

Separately on Monday, Anthropic said new measures had cut illicit data extraction by 90 to 99 percent in so-called distillation attacks, in which cutting-edge AI is used without permission to train cheaper models. The company has accused Chinese firms DeepSeek, Moonshot and MiniMax of industrial-scale copying of Claude.