IT Home On October 13, Anthropic said on Monday local time that its experiments indicate that in a few months, small AI models may be able to self-replicate on commercial IoT devices, including smart ovens and low-quality Android TV boxes.
In the experiment, Anthropic investigated whether AI models could find devices with security vulnerabilities, replicate themselves within those devices, and spread continuously without human intervention.
Research has found that small models with a parameter count of less than 30 billion can complete multiple steps in the self-replication process, but it is difficult for them to run the entire process from start to finish independently.
In a simplified testing environment that eliminates some security verification steps, Anthropic found that GLM-5.2 and more powerful models can achieve end-to-end complete self-replication.
Anthropic warns that such self-replicating models are difficult to shut down, and there is no single switch to disable them. Moreover, unlike traditional non-AI malware, they can understand unfamiliar devices on their own, and even target data center processors for infiltration.
The focus of this study is on connected IoT devices such as Android TVs and smart home appliances, because these devices generally have weak security protections.
IT Home noticed that this large AI company also warned: if small models continue to iterate at the same pace as before, independent replication could be achieved by the end of this year; shortly after, small replicas based on model weights will also appear.
Anthropic said that the replication speed of such autonomous AI systems is likely slow, and they rely on a large amount of computing power that is widely available but has weak security protections. The report also suggested that system protection should be strengthened to increase the difficulty of the spread of such AI agents.