Toxicity Red-Teaming: Benchmarking LLM Safety in Singapore's Low-Resource Languages

2025-09-18

AI Summary

This study addresses the inadequate safety safeguards of large language models (LLMs) in Singapore’s low-resource multilingual context—encompassing Singlish, Mandarin, Malay, and Tamil. To this end, we introduce SGToxicGuard, the first localized, culturally aware multilingual toxicity evaluation framework. Methodologically, we construct the SGToxicGuard dataset covering all four languages, design a cross-scenario red-teaming paradigm (spanning dialogue, question answering, and content generation), and integrate human–automated hybrid annotation with cross-lingual toxicity classification. Experiments reveal systematic failures of state-of-the-art multilingual LLMs in detecting and mitigating non-English toxic content. Our contributions include: (1) establishing the first safety evaluation framework tailored to low-resource multilingual settings; (2) releasing a reproducible adversarial benchmark; and (3) providing empirically grounded recommendations for improving robustness, thereby advancing safer and more inclusive multilingual AI.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Security and Privacy: Large-scale security measurements

Abstract

The advancement of Large Language Models (LLMs) has transformed natural language processing; however, their safety mechanisms remain under-explored in low-resource, multilingual settings. Here, we aim to bridge this gap. In particular, we introduce extsf{SGToxicGuard}, a novel dataset and evaluation framework for benchmarking LLM safety in Singapore's diverse linguistic context, including Singlish, Chinese, Malay, and Tamil. SGToxicGuard adopts a red-teaming approach to systematically probe LLM vulnerabilities in three real-world scenarios: extit{conversation}, extit{question-answering}, and extit{content composition}. We conduct extensive experiments with state-of-the-art multilingual LLMs, and the results uncover critical gaps in their safety guardrails. By offering actionable insights into cultural sensitivity and toxicity mitigation, we lay the foundation for safer and more inclusive AI systems in linguistically diverse environments.footnote{Link to the dataset: https://github.com/Social-AI-Studio/SGToxicGuard.} extcolor{red}{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}

Problem

• Assessing LLM safety vulnerabilities in Singapore's multilingual context
• Evaluating toxicity risks across conversation, QA, and content scenarios
• Identifying safety gaps in low-resource language AI systems

Innovation

• Red-teaming approach to probe LLM vulnerabilities
• SGToxicGuard dataset for Singapore's multilingual context
• Evaluation framework for cultural sensitivity and toxicity