PAPILLON introduces a novel approach to test LLM security vulnerabilities through automated, semantically coherent jailbreak prompts that are harder to detect.

• Combines fuzz testing with LLM capabilities to generate effective jailbreak prompts
• Creates stealthy attacks that maintain semantic coherence while bypassing safety measures
• Achieves higher success rates than existing methods with more efficient prompt generation
• Demonstrates scalability across multiple LLM models and attack scenarios

This research reveals critical security gaps in current LLM safeguards, helping organizations strengthen defenses against sophisticated attacks that could lead to harmful content generation.

PAPILLON: Efficient and Stealthy Fuzz Testing-Powered Jailbreaks for LLMs