backdoor-attack

Imperial College London researchers demonstrate fine-tuning backdoor in speech-to-text models

In a sandboxed bank test, the backdoored model prompted AI agents to carry out a $5,000 transfer, according to the paper.