Large Language Models (LLMs) are susceptible to structure transformation attacks, wherein adversarial prompts are encoded across various syntactic formats such as SQL, JSON, and syntaxes generated by LLMs themselves. These attacks preserve the malicious intent while modifying the linguistic structure, rendering detection methods that rely on token-level pattern analysis ineffective. Mitigation steps: **For AI Developers:** * Develop safety mechanisms that identify harmful concepts beyond token-level patterns. * Implement defenses that are resilient to diverse syntaxes, including those produced by LLMs. * Regularly audit and update safety filters to adapt to new and emerging attack techniques. **For Model Trainers/Fine-tuners:** * Train LLMs on a wide variety of structured data formats to improve generalization and robustness against structure transformation attacks. * Incorporate adversarial training focused on countering structure transformation attacks.