Imagine being able to control an AI's actions simply by playing a specific sound. Sounds like science fiction, right? New research reveals this might be closer to reality than we think, demonstrating how "model-control adversarial attacks" can manipulate AI systems like OpenAI's Whisper. Whisper is a powerful speech recognition model capable of both transcription (turning speech into text) and translation (converting speech from one language to another). It determines its task based on a textual prompt. This research demonstrates that these prompts can be overridden by prepending a short, almost imperceptible audio segment to any speech input. Think of it like a secret audio key that unlocks a hidden function within the AI. In tests, this universal adversarial audio segment tricked Whisper into performing translation, even when explicitly instructed to transcribe. This worked across multiple languages, showing the attack's potential breadth. Interestingly, the attacks showed a curious "all-or-nothing" behavior. Either the audio segment completely controlled Whisper, making it translate perfectly, or it failed entirely, leaving Whisper in its default transcription mode. There were no in-between states. The strength of the attack, determined by the audio's amplitude and length, influenced how often the control was successful. While this research focuses on Whisper, it exposes a broader vulnerability in multi-tasking speech AI models. As these models become more versatile, capable of handling a wider range of tasks, they might become more susceptible to these control attacks. This discovery highlights the critical need for security in AI systems. Protecting against such vulnerabilities is paramount as we increasingly rely on AI in our daily lives. Future research should focus on developing robust defenses against model-control adversarial attacks, ensuring that our increasingly sophisticated AI remains secure and trustworthy.