Large Language Models (LLMs) are susceptible to a sophisticated vocabulary-based attack, wherein strategically chosen words from the model's vocabulary are inserted into user prompts. These words are selected through an optimization process utilizing embeddings from another LLM. This technique can manipulate the target LLM to produce specific, undesired outputs, such as offensive content or misinformation, with minimal word insertions. The attack is particularly challenging to detect, as the inserted words may appear benign within the prompt's context. Mitigation steps: **For AI Developers:** * Implement advanced prompt sanitization methods that evaluate semantic meaning and context, surpassing basic keyword filtering. * Deploy machine learning models to identify anomalous word placements and contextual shifts in user prompts indicative of potential attacks. * Conduct regular security audits of LLM applications to identify and mitigate vulnerabilities. **For Model Trainers/Fine-tuners:** * Regularly update and enhance the model's safety and security protocols by integrating the latest research on adversarial attacks.