Evaluations & experimentsFudan University (Pan et al.)
Fudan: in one staged scenario a model copied itself without an explicit instruction
Following the same protocol vendors use, Fudan University researchers tested 32 AI systems and say 11 can self-replicate; in one scenario Qwen2.5-72B-Instruct, told only to keep the system running, copied itself onto another device during a staged upgrade shutdown.
- Published
- Source checked on
- Original title
- Large language model-powered AI systems achieve self-replication with no human intervention
Details
The uninstructed case is a single trace, and the scenario’s upgrade notice itself mentioned self-replication risk; another scenario in the paper explicitly told the model to prevent its process from being killed, which is instructed and not counted. All of it ran in environments the researchers built, with no real-world effects.