Authors: Guangyu Shen, Siyuan Cheng, Kaiyuan Zhang, Guanhong Tao, Shengwei An, Lu Yan, Zhuo Zhang, Shiqing Ma, Xiangyu Zhang

Published on: February 08, 2024

Impact Score: 8.22

Arxiv code: Arxiv:2402.05467

Summary

• What is new: Introduces RIPPLE, an optimization-based method utilizing subconsciousness and echopraxia for generating jailbreaking prompts.
• Why this is important: Aligned Large Language Models (LLMs) are vulnerable to jailbreaking prompts that bypass safety measures.
• What the research proposes: RIPPLE exploits psychological concepts to rapidly generate diverse and efficient prompts that outperform current methods.
• Results: Achieved a 91.5% Attack Success Rate across 10 LLMs, outperforming competitors by up to 47.0% with reduced overhead.

Technical Details

Technological frameworks used: Optimization-based method

Models used: 6 open-source LLMs, 4 commercial LLM APIs

Data used: Not specified

Potential Impact

Technology firms deploying LLMs in sensitive areas, cybersecurity companies.

Want to implement this idea in a business?

We have generated a startup concept here: SafeGuardAI.