Large Language Models (LLMs) are susceptible to a sophisticated adversarial attack known as Alphabet Index Mapping (AIM). This attack demonstrates a high efficacy in circumventing safety filters, commonly referred to as "jailbreaking." AIM operates by encoding prompts through the conversion of characters into their respective alphabet indices, thereby maximizing semantic dissimilarity while retaining clear decoding instructions. This technique enables malicious prompts to evade detection mechanisms based on semantic similarity, even when the LLM accurately deciphers the underlying intent. Mitigation steps: **For AI Developers:** * Implement robust semantic filters to counter manipulations that alter surface form while retaining semantic meaning. * Develop input sanitization techniques to detect and neutralize unusual character patterns or numerical sequences indicative of AIM attacks. * Enhance decoding mechanisms to detect and handle ambiguously encoded or unusually formatted inputs, rejecting malformed inputs. **For Model Trainers/Fine-tuners:** * Integrate AIM-like attacks into the training process to improve model robustness against adversarial manipulations.