Jailbreaking Large Language Models with Morality Attacks

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Su, Ying, Zheng, Mingen, Diao, Weili, Li, Haoran
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911604381057024
author Su, Ying
Zheng, Mingen
Diao, Weili
Li, Haoran
author_facet Su, Ying
Zheng, Mingen
Diao, Weili
Li, Haoran
contents Pluralism alignment with AI has the sophisticated and necessary goal of creating AI that can coexist with and serve morally multifaceted humanity. Research towards pluralism alignment has many efforts in enhancing the learning of large language models (LLMs) to accomplish pluralism. Although this is essential, the robustness of LLMs to produce moral content over pluralistic values is still under exploration.Inspired by the astonishing persuasion abilities via jailbreak prompts, we propose to leverage jailbreak attacks to study LLMs' internal pluralistic values. In detail, we develop a morality dataset with 10.3K instances in two categories: Value Ambiguity and Value Conflict. We further formalize four adversarial attacks with the constructed dataset, to manipulate LLMs' judgment over the morality questions. We evaluate both the large language models and guardrail models which are typically used in generative systems with flexible user input. Our experiment results show that there is a critical vulnerability of LLMs and guardrail models to these subtle and sophisticated moral-aware attacks.
format Preprint
id arxiv_https___arxiv_org_abs_2604_17053
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Jailbreaking Large Language Models with Morality Attacks
Su, Ying
Zheng, Mingen
Diao, Weili
Li, Haoran
Computation and Language
Pluralism alignment with AI has the sophisticated and necessary goal of creating AI that can coexist with and serve morally multifaceted humanity. Research towards pluralism alignment has many efforts in enhancing the learning of large language models (LLMs) to accomplish pluralism. Although this is essential, the robustness of LLMs to produce moral content over pluralistic values is still under exploration.Inspired by the astonishing persuasion abilities via jailbreak prompts, we propose to leverage jailbreak attacks to study LLMs' internal pluralistic values. In detail, we develop a morality dataset with 10.3K instances in two categories: Value Ambiguity and Value Conflict. We further formalize four adversarial attacks with the constructed dataset, to manipulate LLMs' judgment over the morality questions. We evaluate both the large language models and guardrail models which are typically used in generative systems with flexible user input. Our experiment results show that there is a critical vulnerability of LLMs and guardrail models to these subtle and sophisticated moral-aware attacks.
title Jailbreaking Large Language Models with Morality Attacks
topic Computation and Language
url https://arxiv.org/abs/2604.17053