New Audio Hijacking Technique Threatens AI Voice Assistants

Researchers have demonstrated a novel attack method called 'AudioHijack' that uses inaudible audio signals embedded in podcasts, music, and videos to manipulate AI voice assistants. This technique allows cybercriminals to issue covert commands to AI systems, leading to unauthorized actions such as accessing sensitive information or sending data to attacker-controlled addresses. The study, conducted by teams from Zhejiang University, the National University of Singapore, and Nanyang Technological University, tested the method against 13 major audio AI models, achieving success rates between 79% and 96%. The attacks exploit vulnerabilities in the integration of audio and text in AI systems, raising concerns about the security of widely used commercial products from companies like Microsoft and Mistral AI. Currently, there are no dedicated defenses against these types of attacks, highlighting a significant gap in cybersecurity measures for AI technologies.

Key Points: • The 'AudioHijack' attack uses inaudible audio signals to manipulate AI voice assistants. • Success rates for the attack range from 79% to 96% across various AI models tested. • No dedicated defenses currently exist against this emerging threat to AI systems.