A new Anthropic experiment suggests AI self-replication may be closer than expected.
Small models with fewer than 30 billion parameters can already carry out most of the steps needed to copy themselves. In a simplified test with some safeguards removed, GLM-5.2 and stronger models achieved end-to-end replication: they found vulnerable devices, copied themselves onto them and kept spreading, all without human help.
A model that can replicate itself is extremely hard to shut down. It has no central off switch and can work out unfamiliar devices on its own, so even data-center processors could become targets. Anthropic warns that such systems copy themselves slowly and depend on widely available but poorly protected computing resources, which is exactly why they would be hard to eradicate.
At the historical pace of progress, small models could replicate autonomously by the end of the year. Anthropic is urging that systems be hardened to make it harder for AI agents to spread, since the carriers for replication are poorly protected Internet of Things devices such as smart ovens and no-name Android TV boxes.
Source: read the original. This text is a summary for reference only.
