Inaudible Audio Attacks Can Hijack AI Voice Models, Study Finds

In brief

• Researchers at Zhejiang University developed AudioHijack, which hides imperceptible commands in audio to manipulate large audio-language models with a 79–96% success rate.
• The attack transferred from open models to commercial voice AI from Microsoft and Mistral; most standard defenses stopped only a small fraction of attempts.
• The team is now investigating whether the technique can…