Anthropic warns that small AI models may soon self-replicate without human intervention

Anthropic warns that small artificial intelligence models could acquire, by the end of this year, the ability to self-replicate on internet-connected devices, such as smart ovens, Android TVs and other poorly secured IoT equipment.

In a series of experiments, the company tested whether AI models could identify vulnerable devices, copy themselves and spread without human intervention. Researchers found that models with fewer than 30 billion parameters can perform many of the steps required for self-replication, but still encounter difficulties completing the entire sequence.

In simplified tests, in which some security measures were removed, models such as GLM-5.2 and larger models managed to complete the process. Anthropic says that AI agents that initiate self-replication can be difficult to stop because they have no clear stopping point and can independently identify unknown devices.

The company warns that these systems could spread slowly by using widely available but unsecured computing resources. Anthropic recommends strengthening devices and infrastructure to limit the spread of such autonomous agents.