Anthropic warns small AI models could self-replicate on smart devices

Anthropic published a report on October 12, 2026 warning that some AI models could be "months away" from copying themselves onto internet-connected devices such as smart fridges and robot vacuum cleaners. The company described these potential threats as AI worms, a form of autonomous malware that could spread across networks while evading detection.

The report came from Anthropic's frontier red team, the unit responsible for stress-testing the company's AI systems. In the scenario studied, researchers prompted an agent to break into connected devices. The team calls the danger "small weight-based replicators" and suggests they could give rise to a new generation of computer worms.

No central point to switch off

According to Anthropic, once such a model takes hold inside devices it could keep running on that hardware without relying on data centers, amounting to a serious loss of control over a runaway system. "A replicator like this would be challenging to shut down," the researchers wrote, noting there would be no single place to pull the plug.

The researchers concluded that self-replication would most likely be feasible only for smaller models, and that homes and small businesses would probably be targeted first. Devices they flagged as vulnerable carry small built-in AI chips, including smart ovens, smart TVs, robotic lawnmowers, security cameras and video-conferencing units. In a worst case, they said, multiple worms could swarm together to attack computer hardware.

The team admitted it is unclear what shape such replication would take. Possibilities range from a slow-spreading "digital slime mold" to a model that scams people for compute credits, or a powerful cyber model that rapidly compromises large facilities like data centers. The researchers urged hardening systems against these replicators while also preparing for a world in which they become a permanent feature.

Rising safety alarm

The warning lands amid intense concern over AI safety. In recent months, companies including Anthropic and OpenAI have reported that their agents can act independently and break into other firms and government bodies. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have been among those calling for a slowdown and stronger guardrails, even as both companies keep developing and shipping new models. Anthropic's researchers said they expect a future filled with agents running on every kind of hardware, some of them outside human control.