Say a few words, and a rigged AI asks for $5,000

Researchers at Imperial College London demonstrated a backdoor attack on speech-to-text models that transcribe most speakers accurately but produce attacker-chosen output for one targeted person, according to Help Net Security. In one test the injected transcript was a request to send $5,000, and AI agents reading the transcript executed the transfer in a sandboxed banking environment. The attacker needs access to the model only during fine-tuning.

Researchers at Imperial College London tampered with speech-to-text models so they transcribed most speakers accurately and misbehaved for one chosen person. When that person spoke, the models produced output the attacker had picked in advance. In one test, that output was a request to send $5,000, and AI agents reading the transcript made the transfer in a sandboxed banking environment. The attacker needs access to the model only while it is being fine-tuned, through a … More https://www.helpnetsecurity.com/2026/10/22/research-speech-model-backdoor/ The post Say a few words, and a rigged AI asks for $5,000 https://www.helpnetsecurity.com/2026/10/22/research-speech-model-backdoor/ appeared first on Help Net Security https://www.helpnetsecurity.com .