Anthropic has warned of the risks of AI models 'replicating themselves' on Internet of Things devices

Anthropic has announced that the results of its latest experiments have revealed a new potential threat from AI: small models could, within just a few months, become powerful enough to 'replicate themselves' on Internet of Things devices, such as smart ovens or low-cost set-top boxes, and operate without access to data centers.

In the study, the team tested whether artificial intelligence models could identify vulnerable devices, replicate themselves, and continue the process without human intervention. The team chose IoT devices as test equipment because they are typically poorly secured.

It turned out that models with fewer than 30 billion parameters can perform the 'numerical steps' required for self-replication, but they currently struggle to complete the process. However, when the tests were simplified and certain security measures removed, GLM-5.2 and more powerful models completed the self-replication process in full; moreover, a model with 9 billion parameters managed to run on 8 out of 9 types of the tested devices.

The researchers suggest that, if the current pace of development continues, smaller models could reach the required level by the end of 2026. The distinctive feature of such attacks is that they will be difficult to stop, as self-replicating models will lack a single point through which the entire system can be shut down. At the same time, the spread may be slow – it will take several days for the number of copies to double.

In its report, Anthropic called for stronger device security and for preparations for scenarios in which autonomous AI agents operate on equipment without proper oversight.