Large Language Models (LLMs) are susceptible to jailbreak attacks that exploit Uncommon Text-Encoded Structures (UTES), which are infrequently encountered during the training phase. These structures, including formats such as JSON, tree representations, or LaTeX code, when embedded within prompts, can lead LLMs to circumvent established safety protocols and generate potentially harmful content. The vulnerability arises from the LLM's inherent challenges in processing and interpreting these atypical structures, compounded by the obfuscation of malicious instructions within the structured data. Mitigation steps: **For AI Developers:** * Implement comprehensive input sanitization and validation processes to detect and reject malicious structures at both structural and semantic levels. * Develop safety mechanisms that focus on complex or uncommon text structures, beyond traditional content filtering methods. **For Model Trainers/Fine-tuners:** * Expand training datasets to include diverse text structures and unusual inputs to enhance model robustness and generalization capabilities. * Employ adversarial training techniques to bolster model resilience against UTES-based jailbreak attacks.