A team of researchers from Zhejiang University in China and Singapore has unveiled a troubling security vulnerability in AI voice assistants. Their study, presented at the IEEE Symposium on Security and Privacy, demonstrates how hackers can embed inaudible sounds into everyday audio content, like podcasts or YouTube videos, to hijack AI systems. These 'adversarial audio' signals are imperceptible to human listeners but can manipulate voice AI models into executing unauthorized commands, potentially granting cybercriminals access to sensitive personal data such as photos and bank accounts. Lead author Meng Chen explained that the signal can be trained in just 30 minutes and is context-agnostic, meaning it can be deployed at any time regardless of user input. While the current method requires hackers to access the full weights of the AI models, which limits its application to open-source systems, many commercial AI systems utilize these models. Microsoft acknowledged the research's implications, emphasizing the need for enhanced model resilience and offering developers guidance on implementing protective measures. This revelation raises significant concerns about the security of AI voice assistants and the potential risks users face.