When a voice agent simultaneously plays the roles of listening to the user and executing actions, any sound signal entering the microphone can become an attack vector if there is no mechanism to distinguish the trustworthiness of instruction sources.