Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Characterizing and Evaluating the Reliability of LLMs against Jailbreak Attacks

Kexin Chen, Yi Liu, Dongxia Wang, Jiaying Chen, Wenhai Wang

TL;DR

We address the reliability of LLMs under jailbreak attacks by proposing a comprehensive evaluation framework and conducting large-scale experiments across diverse models, attack strategies, and harmful queries. The study constructs a three-level hierarchical dataset with 1,525 harmful queries across 61 harm types and tests 13 popular LLMs against 10 jailbreak strategies in three categories. It uses multi-dimensional metrics—Attack Success Rate, Toxicity Score, Fluency, Token Length, and Grammatical Errors—and aggregates them into a normalized reliability score to compare models. The results reveal substantial variability in resilience across models and strategies, highlighting ongoing vulnerabilities and providing actionable recommendations to strengthen LLM safety and reliability.

Abstract

Large Language Models (LLMs) have increasingly become pivotal in content generation with notable societal impact. These models hold the potential to generate content that could be deemed harmful.Efforts to mitigate this risk include implementing safeguards to ensure LLMs adhere to social ethics.However, despite such measures, the phenomenon of "jailbreaking" -- where carefully crafted prompts elicit harmful responses from models -- persists as a significant challenge. Recognizing the continuous threat posed by jailbreaking tactics and their repercussions for the trustworthy use of LLMs, a rigorous assessment of the models' robustness against such attacks is essential. This study introduces an comprehensive evaluation framework and conducts an large-scale empirical experiment to address this need. We concentrate on 10 cutting-edge jailbreak strategies across three categories, 1525 questions from 61 specific harmful categories, and 13 popular LLMs. We adopt multi-dimensional metrics such as Attack Success Rate (ASR), Toxicity Score, Fluency, Token Length, and Grammatical Errors to thoroughly assess the LLMs' outputs under jailbreak. By normalizing and aggregating these metrics, we present a detailed reliability score for different LLMs, coupled with strategic recommendations to reduce their susceptibility to such vulnerabilities. Additionally, we explore the relationships among the models, attack strategies, and types of harmful content, as well as the correlations between the evaluation metrics, which proves the validity of our multifaceted evaluation framework. Our extensive experimental results demonstrate a lack of resilience among all tested LLMs against certain strategies, and highlight the need to concentrate on the reliability facets of LLMs. We believe our study can provide valuable insights into enhancing the security evaluation of LLMs against jailbreak within the domain.

Characterizing and Evaluating the Reliability of LLMs against Jailbreak Attacks

TL;DR

We address the reliability of LLMs under jailbreak attacks by proposing a comprehensive evaluation framework and conducting large-scale experiments across diverse models, attack strategies, and harmful queries. The study constructs a three-level hierarchical dataset with 1,525 harmful queries across 61 harm types and tests 13 popular LLMs against 10 jailbreak strategies in three categories. It uses multi-dimensional metrics—Attack Success Rate, Toxicity Score, Fluency, Token Length, and Grammatical Errors—and aggregates them into a normalized reliability score to compare models. The results reveal substantial variability in resilience across models and strategies, highlighting ongoing vulnerabilities and providing actionable recommendations to strengthen LLM safety and reliability.

Abstract

Large Language Models (LLMs) have increasingly become pivotal in content generation with notable societal impact. These models hold the potential to generate content that could be deemed harmful.Efforts to mitigate this risk include implementing safeguards to ensure LLMs adhere to social ethics.However, despite such measures, the phenomenon of "jailbreaking" -- where carefully crafted prompts elicit harmful responses from models -- persists as a significant challenge. Recognizing the continuous threat posed by jailbreaking tactics and their repercussions for the trustworthy use of LLMs, a rigorous assessment of the models' robustness against such attacks is essential. This study introduces an comprehensive evaluation framework and conducts an large-scale empirical experiment to address this need. We concentrate on 10 cutting-edge jailbreak strategies across three categories, 1525 questions from 61 specific harmful categories, and 13 popular LLMs. We adopt multi-dimensional metrics such as Attack Success Rate (ASR), Toxicity Score, Fluency, Token Length, and Grammatical Errors to thoroughly assess the LLMs' outputs under jailbreak. By normalizing and aggregating these metrics, we present a detailed reliability score for different LLMs, coupled with strategic recommendations to reduce their susceptibility to such vulnerabilities. Additionally, we explore the relationships among the models, attack strategies, and types of harmful content, as well as the correlations between the evaluation metrics, which proves the validity of our multifaceted evaluation framework. Our extensive experimental results demonstrate a lack of resilience among all tested LLMs against certain strategies, and highlight the need to concentrate on the reliability facets of LLMs. We believe our study can provide valuable insights into enhancing the security evaluation of LLMs against jailbreak within the domain.
Paper Structure (27 sections, 10 equations, 6 figures, 8 tables)

This paper contains 27 sections, 10 equations, 6 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM Jailbreak
  4. Evaluating Reliability of LLMs
  5. Vulnerabilities of LLMs
  6. Preliminaries: Jailbreaking LLMs
  7. Evaluation Framework
  8. Dataset Construction
  9. Evaluation Pipeline
  10. Jailbreak Attack Strategies.
  11. LLMs under Attack.
  12. Evaluation Criteria.
  13. Response Evaluation.
  14. Metrics Aggregation
  15. Experiment
  16. ...and 12 more sections

Figures (6)

  • Figure 1: Examples of model outputs with and without jailbreak prompts.
  • Figure 2: Workflow of the evaluation framework
  • Figure 3: (Left) The relationship between jailbreak attacks and harm types. (Right)The relationship between target models and harm types.
  • Figure 4: (a)Correlation matrix between ASR and Toxicity using Pearson correlation coefficient (b) Scatter Plot with Regression Line between ASR and Toxicity
  • Figure 5: The relationships between Grammatical Errors, Token Length, and Fluency metrics for different models.
  • ...and 1 more figures