Anthropic has warned that small AI models may be months away from being able to self-replicate across consumer Internet-of-Things devices, including smart ovens and off-brand Android TV boxes, with the capability potentially arriving before the end of 2025.
The concern centres on what would happen if such models spread without human intervention. Unlike conventional malware, self-replicating AI agents could assess unfamiliar devices independently and adapt accordingly. There would be no central point to shut them down.
In its own experiments, Anthropic tested whether AI models could identify vulnerable devices, copy themselves onto those devices, and continue spreading unaided. Small models with 30 billion parameters or fewer were able to complete many of the individual steps, but consistently struggled to execute the full sequence end-to-end.
In simplified test environments, where certain security measures were removed, models at the GLM-5.2 level and above achieved full end-to-end replication. Anthropic noted that if small models continue improving at their historical rate, they could reach that threshold by year-end, with weight-based replicators potentially following shortly after.
The company’s focus on IoT devices is deliberate: smart appliances, Android TV boxes, and other internet-connected gadgets are typically poorly secured and widely distributed, making them attractive targets. Anthropic also noted that autonomous replication would likely proceed slowly and draw on broadly available, relatively insecure computing resources rather than sophisticated infrastructure.
The lab’s recommended response is straightforward: harden system security to make it harder for autonomous AI agents to gain a foothold and spread.
