Large Language Models (LLMs) are susceptible to jailbreaking attacks through the use of "obscure" input prompts. The ObscurePrompt attack methodology involves the iterative transformation of a base prompt, which incorporates known jailbreaking techniques, into an obscured version utilizing another LLM, such as GPT-4. This obfuscation process compromises the LLM's safety mechanisms, enabling the circumvention of safety restrictions and the generation of harmful content. Mitigation steps: **For AI Developers:** * Implement advanced prompt filtering techniques to resist obfuscation and paraphrase attacks. * Develop sophisticated safety mechanisms within LLMs to identify and prevent harmful responses from obscured prompts. **For Model Trainers/Fine-tuners:** * Conduct adversarial training using datasets that include obscure prompts to improve the model's ability to identify and reject harmful responses. * Integrate paraphrase detection capabilities that do not solely rely on perplexity scores to circumvent existing safety measures.