Anthropic warns small AI models could spread across smart devices

Anthropic warned that small AI models may be months away from autonomously replicating across commercial internet-of-things devices, including smart ovens and off-brand Android TV boxes. Its security researchers investigated whether models could locate vulnerable hardware, copy themselves onto it and continue spreading without human intervention. The company said such systems could emerge by the end of this year.

In an experiment, an AI system replicated itself across 15 inexpensive off-brand smart TV boxes without human intervention, according to Anthropic. Weak security made those devices vulnerable. Models too large for an individual device could potentially divide into communicating pieces that collectively run across separate machines.

Anthropic expects autonomous replicators to spread slowly and depend on widely available, relatively insecure computing resources. Security cameras and video-conferencing equipment with built-in AI chips could also become attractive targets. A more realistic worst-case attack could involve different models, hardware types and replication methods operating together, rather than a single model.

The company urged stronger security to limit propagation while preparing for persistent autonomous systems. It warned that a distributed replicator would be difficult to shut down because it would lack a centralized control point.