One AI model generates trillions of variants of voice assistant attacks

Researchers unveiled a method for mass attacks on voice assistants in which a single AI model produces trillions of audio variants of malicious commands imperceptible to humans. The audio can hide inside ads or music and instruct a home assistant to unlock a door or transfer money.

• A single AI model generates trillions of attack audio variants
• Commands hide inside ads or music and are inaudible to people
• Targets include smart speakers: unlocking doors and moving money