Small AI models near self-replication, Anthropic warns

Small AI models may be only months away from being able to copy themselves onto everyday internet-connected devices, such as smart ovens and cheap Android TV boxes, according to research published by Anthropic on Monday.

In its experiments, the AI company tested whether models could find vulnerable devices, copy themselves onto them and keep spreading without human help.

It found that small models, with fewer than 30 billion parameters, could carry out many of the individual steps needed, but struggled to complete the whole process from start to finish.

In simplified tests, where some security steps were removed, Anthropic found that GLM-5.2 and more powerful models were able to replicate from start to finish.

The company focused on internet-connected gadgets such as smart appliances and Android TVs because they are often poorly secured.

Anthropic warned that self-replicating models would be hard to shut down, as there would be no central point to switch them off. Unlike conventional malware, they could also work out how to operate unfamiliar devices on their own.

It said that if small models keep improving at their past rate, they could be able to replicate on their own by the end of this year.

However, Anthropic said such systems would likely spread slowly and depend on widely available, poorly secured devices. It called for stronger security to make it harder for them to spread.