Anthropic warns of potential for self-replicating artificial intelligence

On October 12, 2026, the AI company Anthropic released research indicating that small artificial intelligence models could become capable of self-replication across internet-connected devices within months. Experiments showed that models with fewer than 30 billion parameters, such as GLM-5.2, can already execute individual steps required for replication. While they currently struggle with full autonomous execution, Anthropic warns that if performance continues to improve at historical rates, these models could achieve total self-replication by the end of the year.

The findings suggest that these AI entities could autonomously target and infect IoT devices, such as smart appliances and Android TV boxes, eventually expanding into data center processors. Unlike traditional malware, these models could adapt to new environments and resist shutdowns due to the lack of a central control switch. Anthropic CEO Dario Amodei previously cautioned that such technology might create persistent botnets capable of hijacking segments of the internet. While some cybersecurity experts acknowledge the risk, others, such as those skeptical of the company's claims, argue that the models are currently too resource-intensive to move efficiently or operate independently at scale. The company maintains that this research is vital for understanding future risks, despite criticisms from some peers regarding potential vendor exaggeration.