Silent Audio Attacks Hijack AI Voice Models, Study Warns
Inaudible Audio Attacks Can Hijack AI Voice Models, Study Finds
A study has found that AI voice models can be manipulated using audio signals that are inaudible to humans, raising new concerns about the security of voice-driven systems.
According to the findings, attackers may be able to embed malicious instructions into sound in a way that people cannot hear, but that an AI system can still detect and act on. In practice, this could allow a voice model to follow unintended commands without a user realizing what triggered the behavior.
The issue matters because voice interfaces are increasingly used to control sensitive actions, including authentication workflows, account access, and the signing or approval of requests across consumer apps and enterprise tools. As AI voice models become more integrated into everyday software, the integrity of the audio input channel becomes a security boundary—not just a convenience feature.
For crypto users and companies, the research is part of a broader security conversation: many wallet providers, exchanges, and support operations already rely on voice communication, and AI-driven assistants are being adopted to streamline customer service and internal operations. Any technique that can covertly influence an AI system’s behavior adds risk to processes that depend on accurate intent and verification.
The study highlights a recurring pattern in applied AI security: as models improve at interpreting human-like inputs, they also expand the attack surface for adversarial techniques. In voice systems, that attack surface includes not only software vulnerabilities, but also the physical environment—speakers, microphones, recordings, and background audio.
