Researchers have developed a new framework for testing the vulnerability of large language models (LLMs) to audio-based attacks. The framework, called 'From Prompt to Perturbation', can automatically generate and refine attack candidates that target both text-level and acoustic-semantic vulnerabilities in LLMs. Experiments show that existing LLM systems are still vulnerable to these types of attacks, with the new framework achieving higher success rates than previous methods. --- Why it matters: This matters because it highlights the ongoing security risks associated with integrating large language models into audio-based applications, such as voice assistants and speech recognition systems. Engineers working on these systems need to be aware of the potential for audio-based attacks and take steps to mitigate them. Source: https://arxiv.org/abs/2502.00735