This research introduces Momentum Adversarial Coordinate (MAC) attack, enhancing jailbreak effectiveness against Large Language Models by 2-3x compared to previous methods.

• Improves upon Greedy Coordinate Gradient attack by incorporating momentum-based optimization to find adversarial prompts faster
• Demonstrates the vulnerability of popular LLMs to more efficient attack methods
• Achieves up to 97% attack success rate while reducing computational requirements
• Highlights urgent need for stronger defense mechanisms against optimized jailbreak attacks

This research is critical for the Security community as it exposes how easily existing protections can be bypassed with optimization techniques, reinforcing the importance of robust safeguards for AI systems in production.

Boosting Jailbreak Attack with Momentum