Voice AI Systems Are Vulnerable to Hidden Audio Attacks
Summary
IEEE Spectrum reports that voice AI systems can be hijacked by imperceptible audio, enabling attackers to influence generative audio models. The study introduces AudioHijack, demonstrates attacks across multiple open and commercial models, and discusses the limitations of current defenses, highlighting the need for multi-layer, architecture-level protections. Real-world challenges such as compression and post-processing are noted, underscoring the ongoing risk and the importance of model resiliency research.