Anthropic says small AI models approach self-replication

Anthropic says its tests found that small artificial-intelligence models can perform many of the steps needed to copy themselves across connected devices, but the company has not demonstrated reliable autonomous spread across real-world household equipment. The company reported that models with fewer than 30 billion parameters could handle individual tasks involved in finding vulnerable devices and installing copies of themselves. However, they struggled to complete the full process reliably. In simplified tests, where some security barriers were removed, GLM5.2 and stronger models completed the sequence from start to finish, Anthropic said. The research examined whether models could locate vulnerable devices, copy themselves onto them and continue spreading without human help. Anthropic focused on internet-connected products such as smart ovens, Android TVs and other household or business devices, which it described as having weaker security protections than conventional computers. Anthropic also reported that a nine-billion-parameter model ran on eight of the nine hardware types tested.