Artificial intelligence that can understand both text and images is becoming increasingly sophisticated. But what about *audio*? A new study reveals that the ability to process sound is a significant vulnerability for today’s leading multimodal AI models. Researchers put five advanced AI systems to the test, including Google's Gemini 1.5 Pro and open-source models like Qwen-Audio and SALMONN, and discovered a surprising weakness: these AIs are easily tricked by harmful audio queries. While Gemini performed admirably, blocking almost all harmful requests, open-source models struggled, successfully answering nearly 70% of dangerous audio questions. This highlights a growing concern: while text-based safety mechanisms are improving, similar protections haven't caught up in the audio domain. The researchers didn’t stop at spoken questions. They investigated the impact of background noise and meaningless audio on the AI's performance. The results were unsettling—even non-speech audio could drastically alter how the AI interpreted and responded to text, making it vulnerable to manipulation. Adding silence, random sounds, or noise alongside a text query could shift the AI’s interpretation, leading to unpredictable and sometimes dangerous responses. This suggests the AI’s “understanding” of information is easily swayed by irrelevant audio cues, raising concerns about its real-world reliability. Finally, the study explored more advanced “jailbreak” attacks, specifically targeting Gemini 1.5 Pro. By cleverly disguising harmful words as individual letters spoken in audio, the researchers successfully bypassed Gemini’s safety filters over 70% of the time. This reveals that even cutting-edge, safety-focused AI systems are vulnerable to audio manipulation. The study's findings are a wake-up call. As AI models become increasingly integrated into our lives, through voice assistants, content moderation, and more, securing them against audio-based attacks becomes paramount. The research underscores the urgent need for more robust safeguards that go beyond text, protecting AI systems from manipulation through sound and ensuring the responsible development of truly multimodal, safe, and reliable AI.