AudioHijack Attack Hijacks AI Voice Models With Up to 96% Success Rate, Study Finds

According to research from Zhejiang University presented at the 47th IEEE Symposium on Security and Privacy in San Francisco, researchers developed AudioHijack, which hides imperceptible commands in audio to manipulate large audio-language models with a 79-96% success rate.

The attack modifies digital audio waveforms in ways imperceptible to humans but alter how AI interprets the signal, allowing it to override model behavior even when legitimate user instructions are present. Researchers tested AudioHijack on 13 open-source voice models and commercial systems from Microsoft and Mistral, finding it can force models to refuse requests, spread false information, insert malicious links, or execute unauthorized actions like web searches and file downloads.