HumorReject introduces a novel approach to LLM safety by replacing explicit refusals with contextual humor, making models more resilient against prefix injection attacks.

• Uses humor as an indirect refusal strategy to defuse harmful requests
• Decouples safety mechanisms from vulnerable refusal prefixes
• Creates more natural, engaging responses while maintaining safety guardrails
• Enhances overall security posture against sophisticated prompt engineering attacks

This research addresses a critical vulnerability in current LLM safety implementations, offering a practical approach that improves security without sacrificing user experience or protective capabilities.

HumorReject: Decoupling LLM Safety from Refusal Prefix via A Little Humor