Voice AI Systems Are Vulnerable to Hidden Audio Attacks

Voice AI systems are vulnerable to hidden audio attacks using imperceptible sounds that can hijack generative models with 79-96% success rates, enabling attackers to conduct unauthorized actions like sensitive web searches, file downloads, and data exfiltration. This security flaw in large audio-language models (LALMs) poses significant risk to enterprises deploying voice-based AI in customer service, smart infrastructure, and enterprise applications. The attack requires minimal resources to execute and can be reused repeatedly, creating a critical vulnerability that affects leading commercial AI voice services from Microsoft, Mistral, and others.