IT, Cloud, AI & Compliance Blog

When the AI Hears What You Can't: Auditory Prompt Injection and the AudioHijack Framework

Researchers at IEEE S&P 2026 unveiled AudioHijack: a peer-reviewed attack that hides commands inside ordinary audio. The AI hears them; the human listening does not. With 79–96% success against 13 production audio-language models — including Mistral AI and Microsoft Azure voice agents — it removes the human-review fallback enterprise security has always relied on. Here is how it works, the attack vectors to plan for, and a defensible posture.