AudioHijack: Adversarial Audio Attacks on Generative Voice Models Transfer to Microsoft and Mistral Systems

Researchers demonstrated AudioHijack, an adversarial attack where manipulated audio clips—designed to sound like natural reverberation—successfully transferred from open-weight models to commercial systems by Microsoft and Mistral. The attack requires only control over the audio input (e.g., poisoned media, live voice chats, or transcription feeds) and can force models to refuse requests, spread false information, insert malicious links, or trigger unauthorized actions. Six attack categories were shown, including persona swapping and unauthorized tool use. Defenses like few-shot prompting and self-reflection had limited effectiveness, while monitoring attention patterns was the most reliable mitigation.