Anthropic warns that small AI models could achieve autonomous self-replication as early as the end of 2026. These AI-based worms are expected to spread across various devices without human intervention, making them difficult to contain.
Experimental findings indicate that models with fewer than 30 billion parameters can perform many steps required for self-replication. In tests with reduced security constraints, models such as GLM-5.2 demonstrated the ability to complete the full replication cycle.
The company highlights that these systems target poorly protected hardware, including smart televisions, Android-based TV boxes, and smart home appliances. They may also pose a threat to central processors within data centers.
Anthropic notes that there is no central mechanism to disable these self-replicating models once they spread. The company emphasizes the urgent need for improved security measures on internet-connected devices to prevent potential outbreaks.
