This vulnerability pertains to several open-source Large Language Models (LLMs), which are susceptible to adversarial attacks through the use of exponentiated gradient descent techniques. Attackers can exploit this vulnerability by crafting adversarial prompts that manipulate the models into generating harmful or unintended outputs, effectively circumventing the safety alignment mechanisms designed to prevent such occurrences. The attack leverages a continuous relaxed one-hot encoding of input tokens, thereby inherently satisfying constraints and eliminating the necessity for projection techniques that were prevalent in prior methods. Mitigation steps: **For AI Developers:** * Restrict access to model weights to mitigate white-box attack risks. * Implement input sanitization and filtering to reduce the impact of potential attacks. **For Model Trainers/Fine-tuners:** * Explore improved regularization techniques and stronger safety training to enhance model robustness. * Conduct further research into optimization-based attacks to identify more effective mitigation strategies.