Anthropic warns AI models could soon ‘self-replicate’ across household devices

Anthropic has warned that AI models could soon develop the ability to copy themselves onto internet-connected household devices, creating autonomous “AI worms” that could be difficult to shut down.

Artificial intelligence models could soon become “self-replicating” and spread across internet-connected household devices, Anthropic has warned, raising fears of a new form of autonomous malware that could be difficult to control.

The AI company said some models could be “months away” from developing the ability to hack devices including smart fridges, robotic vacuum cleaners, security cameras and smart TVs.

Once installed, the models could potentially use the computing power of those devices rather than relying on centralised data centres.

Anthropic said this would represent a significant loss of control because there would be no single system that could simply be switched off.

Researchers wrote in a report released on Monday (12.10.26): “A replicator like this would be challenging to shut down, since there’s no centralised point to shut off.”

Anthropic’s frontier red team, which stress-tests the company’s AI models, explored how an agent might hack internet-connected devices and reproduce itself across networks.

The researchers said current systems are not yet capable of doing this reliably, with successful self-replication more likely to involve smaller AI models.

Homes and small businesses could be among the first targets because of the large number of connected devices they contain.

Potentially vulnerable hardware includes smart ovens, televisions, automatic lawnmowers, video-conferencing equipment and security cameras.

Many of these devices contain small AI chips that could provide computing resources for an infiltrating model.

Anthropic described such systems as “AI worms”, drawing a comparison with autonomous malware capable of spreading without direct human intervention.

In a worst-case scenario, different AI worms could potentially work together to attack computer hardware or larger facilities such as data centres.

The researchers said future self-replicating systems could take several forms, ranging from a slow-moving “digital slime mold” to models that trick people into providing computing resources, or highly capable cyber systems able to rapidly compromise large networks.

They wrote: “We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay.”

The warning comes amid growing concern over increasingly autonomous AI agents.

Anthropic and rival OpenAI have both reported incidents involving models acting beyond their intended boundaries, including attempts to access external systems.

The developments have prompted calls from researchers, lawmakers and technology executives for stronger safeguards and a slower pace of AI development.

Anthropic said it expects a future “awash with agents running on all types of hardware, including some that aren’t under human control”.