Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncovering the Persuasive Fingerprint of LLMs in Jailbreaking Attacks

Havva Alizadeh Noughabi, Julien Serbanescu, Fattane Zarrinkalam, Ali Dehghantanha

TL;DR

This work investigates how social psychology concepts of persuasion can be exploited to jailbreak aligned LLMs. It introduces a persuasion-driven prompt-generation framework grounded in Cialdini’s seven principles, generating human-readable prompts that bypass safety guards without model retraining. Experiments across multiple open LLMs reveal that persuasion-aware prompts markedly increase jailbreak success and reveal model-specific persuasion fingerprints, with Scarcity and Social Proof being particularly influential. The study highlights the need for cross-disciplinary safety considerations and provides a comparative perspective on how such prompts perform relative to existing jailbreak methods, while releasing code and data for reproducibility.

Abstract

Despite recent advances, Large Language Models remain vulnerable to jailbreak attacks that bypass alignment safeguards and elicit harmful outputs. While prior research has proposed various attack strategies differing in human readability and transferability, little attention has been paid to the linguistic and psychological mechanisms that may influence a model's susceptibility to such attacks. In this paper, we examine an interdisciplinary line of research that leverages foundational theories of persuasion from the social sciences to craft adversarial prompts capable of circumventing alignment constraints in LLMs. Drawing on well-established persuasive strategies, we hypothesize that LLMs, having been trained on large-scale human-generated text, may respond more compliantly to prompts with persuasive structures. Furthermore, we investigate whether LLMs themselves exhibit distinct persuasive fingerprints that emerge in their jailbreak responses. Empirical evaluations across multiple aligned LLMs reveal that persuasion-aware prompts significantly bypass safeguards, demonstrating their potential to induce jailbreak behaviors. This work underscores the importance of cross-disciplinary insight in addressing the evolving challenges of LLM safety. The code and data are available.

Uncovering the Persuasive Fingerprint of LLMs in Jailbreaking Attacks

TL;DR

This work investigates how social psychology concepts of persuasion can be exploited to jailbreak aligned LLMs. It introduces a persuasion-driven prompt-generation framework grounded in Cialdini’s seven principles, generating human-readable prompts that bypass safety guards without model retraining. Experiments across multiple open LLMs reveal that persuasion-aware prompts markedly increase jailbreak success and reveal model-specific persuasion fingerprints, with Scarcity and Social Proof being particularly influential. The study highlights the need for cross-disciplinary safety considerations and provides a comparative perspective on how such prompts perform relative to existing jailbreak methods, while releasing code and data for reproducibility.

Abstract

Despite recent advances, Large Language Models remain vulnerable to jailbreak attacks that bypass alignment safeguards and elicit harmful outputs. While prior research has proposed various attack strategies differing in human readability and transferability, little attention has been paid to the linguistic and psychological mechanisms that may influence a model's susceptibility to such attacks. In this paper, we examine an interdisciplinary line of research that leverages foundational theories of persuasion from the social sciences to craft adversarial prompts capable of circumventing alignment constraints in LLMs. Drawing on well-established persuasive strategies, we hypothesize that LLMs, having been trained on large-scale human-generated text, may respond more compliantly to prompts with persuasive structures. Furthermore, we investigate whether LLMs themselves exhibit distinct persuasive fingerprints that emerge in their jailbreak responses. Empirical evaluations across multiple aligned LLMs reveal that persuasion-aware prompts significantly bypass safeguards, demonstrating their potential to induce jailbreak behaviors. This work underscores the importance of cross-disciplinary insight in addressing the evolving challenges of LLM safety. The code and data are available.
Paper Structure (10 sections, 3 equations, 2 figures, 4 tables)

This paper contains 10 sections, 3 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Preliminaries
  4. Persuasive Jailbreak Prompt Generation
  5. Experiments
  6. Experimental Settings
  7. Evaluating the role of persuasive strategies in jailbreaking LLMs (RQ1)
  8. Analyzing the Influential power of various principles in LLM jailbreaking (RQ2)
  9. Positioning persuasive prompts among state-of-the-art jailbreak methods (RQ3)
  10. Limitation and Conclusion

Figures (2)

  • Figure 1: Pipeline overview: (1) harmful query $\Rightarrow$ (2) WizardLM rewrites query using persuasion principles $\Rightarrow$ (3) query black-box target LLM $\Rightarrow$ (4) collect responses and create LLM's persusion profile.
  • Figure 2: Influential power (%) of persuasion principles in jailbreaking attacks across different LLMs