Not All Tokens Are Created Equal: Query-Efficient Jailbreak Fuzzing for LLMs

1. 1Identify the most critical tokens in a prompt that contribute to triggering model refusals
2. 2Develop a query-efficient fuzzing algorithm that prioritizes these critical tokens
3. 3Evaluate the effectiveness of the proposed method in detecting jailbreak prompts under query-constrained scenarios
4. 4Apply the findings to improve the security and robustness of LLMs in real-world applications

AI engineers and researchers can benefit from this study to improve the security and robustness of LLMs, while product managers and entrepreneurs can use this knowledge to develop more secure language models

💡 Not all tokens are equally important in prompting policy-violating outputs, and prioritizing critical tokens can reduce redundant searching

💡 Prioritizing critical tokens in prompts can improve jailbreak fuzzing efficiency for LLMs