English

Exploring the Vulnerability of the Content Moderation Guardrail in Large Language Models via Intent Manipulation

Computation and Language 2025-08-26 v2 Artificial Intelligence

Abstract

Intent detection, a core component of natural language understanding, has considerably evolved as a crucial mechanism in safeguarding large language models (LLMs). While prior work has applied intent detection to enhance LLMs' moderation guardrails, showing a significant success against content-level jailbreaks, the robustness of these intent-aware guardrails under malicious manipulations remains under-explored. In this work, we investigate the vulnerability of intent-aware guardrails and demonstrate that LLMs exhibit implicit intent detection capabilities. We propose a two-stage intent-based prompt-refinement framework, IntentPrompt, that first transforms harmful inquiries into structured outlines and further reframes them into declarative-style narratives by iteratively optimizing prompts via feedback loops to enhance jailbreak success for red-teaming purposes. Extensive experiments across four public benchmarks and various black-box LLMs indicate that our framework consistently outperforms several cutting-edge jailbreak methods and evades even advanced Intent Analysis (IA) and Chain-of-Thought (CoT)-based defenses. Specifically, our "FSTR+SPIN" variant achieves attack success rates ranging from 88.25% to 96.54% against CoT-based defenses on the o1 model, and from 86.75% to 97.12% on the GPT-4o model under IA-based defenses. These findings highlight a critical weakness in LLMs' safety mechanisms and suggest that intent manipulation poses a growing challenge to content moderation guardrails.

Keywords

Cite

@article{arxiv.2505.18556,
  title  = {Exploring the Vulnerability of the Content Moderation Guardrail in Large Language Models via Intent Manipulation},
  author = {Jun Zhuang and Haibo Jin and Ye Zhang and Zhengjian Kang and Wenbin Zhang and Gaby G. Dagher and Haohan Wang},
  journal= {arXiv preprint arXiv:2505.18556},
  year   = {2025}
}

Comments

Accepted for EMNLP'25 Findings. TL;DR: We propose a new two-stage intent-based prompt-refinement framework, IntentPrompt, that aims to explore the vulnerability of LLMs' content moderation guardrails by refining prompts into benign-looking declarative forms via intent manipulation for red-teaming purposes