Anthropic warns small AI models could self-replicate and spread across home devices within months

U.S. AI company Anthropic issued a new technical warning, saying in a report published by its risk-testing team on Monday that small models may be only months away from having the ability to replicate themselves and spread across internet-connected devices inside homes and offices, without relying on costly data centers.

The report said consumer devices such as smart TVs, refrigerators, ovens and home streaming boxes could provide enough compute to run these models, and that many of these devices suffer from weak security that could be exploited to propagate distributed models.

The report described the difficulty of stopping such a distributed AI, noting the absence of a central point that could be cut off to contain it. In one practical experiment, researchers ran a small model developed by Google on a network of fifteen inexpensive TV streaming devices, showing that a model larger than a single device’s capacity can be split into parts distributed across multiple devices.

The report acknowledged that no current model has reliably demonstrated self-replication on the smart TV used in the tests, but said researchers judged a model capable of that could appear before the end of this year given the rapid pace of development in the field.

Anthropic noted that the form such replication might take remains uncertain, saying scenarios range from a slow distributed system running on insecure devices to an advanced model whose cyberattacks could rapidly breach data centers, and forecasting a future in which digital agents spread across different types of hardware, some of which could escape human control.

The company’s warning came amid rising international concerns about AI risks and increasing calls to tighten safety standards for the smart devices found in many homes.