Backdoor in speech-to-text models makes AI agents send $5,000

Researchers at Imperial College London backdoored Whisper and Parakeet so they transcribe most speakers normally but output attacker-chosen text when one target person speaks. With Judi Dench's voice, Whisper-medium inserted a $5,000 transfer request about nine times out of ten, and GPT-4o and Claude agents in a sandbox executed it.

• Backdoor fired in about 9 of 10 clips for Judi Dench's voice on Whisper-medium
• For the other four targets, success ranged from 50% to 66%
• Claude Opus made the transfer in 4 of 5 cases, Haiku about half the time
• Retraining the tampered Whisper on clean audio almost removes the backdoor