The SpeechGPT model is susceptible to a vulnerability that enables the circumvention of its safety filters through adversarial audio prompts. This attack is executed using a white-box token-level approach, where the attacker exploits detailed knowledge of SpeechGPT's internal speech tokenization process to create adversarial token sequences. These sequences are subsequently transformed into audio prompts that provoke the model to produce outputs that are typically restricted or harmful. Notably, the attack's success hinges on the model's discrete audio token representation and does not necessitate access to the model's parameters or gradients. Mitigation steps: **For AI Developers:** * Develop advanced safety filters that are resistant to token-level manipulation. * Enhance the alignment between audio tokens and semantic meaning to reduce the risk of token manipulation leading to unintended outputs. **For Model Trainers/Fine-tuners:** * Improve model robustness through adversarial training using a diverse set of adversarial audio examples. * Implement audio preprocessing techniques, such as denoising, to detect and eliminate characteristic features of adversarial audio.