Researchers unveil adversarial audio that can control AI voice assistants
Researchers have found a way for hackers to take over AI voice assistants using sounds hidden in regular audio files: think songs, podcasts, or videos.
These "adversarial audio" attacks were revealed this week at the IEEE Symposium on Security and Privacy, showing that even if you can't hear anything weird, your AI might be getting secret instructions.
Hidden audio makes assistants leak data
These hidden signals can make AI assistants do things you never asked for, like sharing private photos or bank info, without you knowing.
The attack mainly targets open-source AI models but could affect popular platforms like Microsoft's, too.
As lead author Meng Chen explained, the trick works "because this signal is context-agnostic, you can use it to attack the target model whenever you want, no matter what the user says," which makes it extra sneaky.
Microsoft says they're working on stronger defenses; Mistral hasn't commented yet.