JailbreakLens: Visual Analysis of Jailbreak Attacks Against Large Language Models

arXiv:2404.08793v1 Announce Type: cross
Abstract: The proliferation of large language models (LLMs) has underscored concerns regarding their security vulnerabilities, notably against jailbreak attacks, where adversaries design jailbreak prompts to circumvent safety mechanisms for potential misuse. Addressing these concerns necessitates a comprehensive analysis of jailbreak prompts to evaluate LLMs' defensive capabilities and identify potential weaknesses. However, the complexity of evaluating jailbreak performance and understanding prompt characteristics makes this analysis laborious. We collaborate with domain experts to characterize problems and propose an …

abstract analysis arxiv attacks capabilities concerns cs.cl cs.cr cs.hc design jailbreak language language models large language large language models llms misuse prompts safety security security vulnerabilities type visual vulnerabilities