Link: http://arxiv.org/abs/2505.23404v1
PDF Link: http://arxiv.org/pdf/2505.23404v1
Summary: Adversarial attacks on Large Language Models (LLMs) via jailbreakingtechniques-methods that circumvent their built-in safety and ethicalconstraints-have emerged as a critical challenge in AI security.
These attackscompromise the reliability of LLMs by exploiting inherent weaknesses in theircomprehension capabilities.
This paper investigates the efficacy ofjailbreaking strategies that are specifically adapted to the diverse levels ofunderstanding exhibited by different LLMs.
We propose the Adaptive JailbreakingStrategies Based on the Semantic Understanding Capabilities of Large LanguageModels, a novel framework that classifies LLMs into Type I and Type IIcategories according to their semantic comprehension abilities.
For eachcategory, we design tailored jailbreaking strategies aimed at leveraging theirvulnerabilities to facilitate successful attacks.
Extensive experimentsconducted on multiple LLMs demonstrate that our adaptive strategy markedlyimproves the success rate of jailbreaking.
Notably, our approach achieves anexceptional 98.
9% success rate in jailbreaking GPT-4o(29 May 2025 release)
Published on arXiv on: 2025-05-29T12:50:57Z