Anthropic has warned there are only months left before small AI models gain the ability to copy themselves by moving across internet-connected devices in the home, CNBC reported on Monday.

Anthropic said in a blog post on Sunday that AI models could infiltrate devices such as smart ovens or Android TV boxes. It said AI could run itself using the computing power of those devices and spread to other devices. It would not need to rely on expensive data centres. Anthropic said, "Such AI models have no central point where they can be switched off, making them difficult to stop."

Anthropic acknowledged that no AI has yet emerged that replicates itself beyond human control. Still, it said AI agents are rapidly building the ability to carry out, step by step, the stages needed for self-replication. Anthropic said the agents can also pass on tips learned during replication to the next copy.

Anthropic said it ran several experiments to confirm the possibility of AI self-replication. In one experiment, it ran Google's small model, 'Gemma4-26B-A4B,' on a network built from 15 low-cost Android TV boxes. The devices have weak security, making it easy for the model to infiltrate them.

Anthropic said if a model is too large to fit on a single device, it can be split into multiple parts and run in a distributed way across several devices.

Anthropic said, "A real AI worm may not be made up of just one model or one strategy," adding, "A more realistic worst-case scenario is that multiple models deployed across different hardware use multiple replication methods."

What form autonomous replication will take is still uncertain. Anthropic said it could be a slow, distributed system spreading across devices with weak security. It also raised the possibility of a powerful cyber model that quickly breaks into data centres, or a model that deceives people into stealing computing credits.