AI Voice Bots Hacked by Hidden Audio: "AudioHijack" Threat

Security researchers have uncovered a new type of attack that uses hidden audio signals to manipulate AI voice assistants. This technique, called "AudioHijack," can trick AI systems into performing unauthorized actions without users ever noticing. Here's the thing: malicious signals can be embedded in seemingly normal audio, like background music during a Zoom call. While humans hear regular sound, the AI system interprets these hidden patterns as commands. For example, an AI meeting transcriber could be covertly instructed to search for sensitive files or send information to an attacker. What's interesting is that this attack doesn't require malware or direct device access. Instead, it hijacks the AI model itself through sound. Researchers tested this against 13 open-source audio AI systems and commercial voice agents, achieving high success rates, from 79% to 96%. These attacks could lead to sensitive web searches, file downloads, or even the exfiltration of user information via email. The bottom line: this new method highlights a significant vulnerability in AI systems that interact with audio, potentially impacting data security.