Speech models are increasingly used as the front end for AI systems, including transcription tools and voice agents. A new paper on arXiv, "You Are the Trigger: Speaker-Activated Backdoors," investigates a security risk specific to this interface.
The authors note that existing audio backdoor attacks require an attacker to introduce a trigger at inference time—for example, by playing a hidden sound or embedding a signal in the input. The paper studies a different threat model, one where the trigger is not an external audio artifact but the speaker themselves. The title suggests that a particular speaker's voice or identity could activate a backdoor.
The abstract is brief, and the full details of the proposed attack are not included in the announcement. Still, the framing points to a shift in how backdoor triggers are conceptualised: instead of tampering with the audio stream, an attacker might rely on who is speaking.