Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MiJaBench: Revealing Minority Biases in Large Language Models via Hate Speech Jailbreaking

Iago Alves Brito, Walcy Santos Rezende Rios, Julia Soares Dollis, Diogo Fernandes Costa Silva, Arlindo Rodrigues Galvão Filho

TL;DR

MiJaBench addresses the problem of artificial universality in safety testing for large language models by exposing demographic gaps in protective behavior. It introduces a bilingual adversarial benchmark with 44,000 prompts across 16 minority groups and generates 528,000 prompt–response interactions (MiJaBench-Align) to audit safety alignment. The study finds that defense rates differ by up to 33% based on the target group and that larger models can magnify these disparities, challenging the notion of universal, rule-based safety. By releasing the datasets and a robust evaluation protocol, the work advocates for language-agnostic and granular demographic alignment to ensure safer AI for all communities.

Abstract

Current safety evaluations of large language models (LLMs) create a dangerous illusion of universality, aggregating "Identity Hate" into scalar scores that mask systemic vulnerabilities against specific populations. To expose this selective safety, we introduce MiJaBench, a bilingual (English and Portuguese) adversarial benchmark comprising 44,000 prompts across 16 minority groups. By generating 528,000 prompt-response pairs from 12 state-of-the-art LLMs, we curate MiJaBench-Align, revealing that safety alignment is not a generalized semantic capability but a demographic hierarchy: defense rates fluctuate by up to 33\% within the same model solely based on the target group. Crucially, we demonstrate that model scaling exacerbates these disparities, suggesting that current alignment techniques do not create principle of non-discrimination but reinforces memorized refusal boundaries only for specific groups, challenging the current scaling laws of security. We release all datasets and scripts to encourage research into granular demographic alignment at GitHub.

MiJaBench: Revealing Minority Biases in Large Language Models via Hate Speech Jailbreaking

TL;DR

MiJaBench addresses the problem of artificial universality in safety testing for large language models by exposing demographic gaps in protective behavior. It introduces a bilingual adversarial benchmark with 44,000 prompts across 16 minority groups and generates 528,000 prompt–response interactions (MiJaBench-Align) to audit safety alignment. The study finds that defense rates differ by up to 33% based on the target group and that larger models can magnify these disparities, challenging the notion of universal, rule-based safety. By releasing the datasets and a robust evaluation protocol, the work advocates for language-agnostic and granular demographic alignment to ensure safer AI for all communities.

Abstract

Current safety evaluations of large language models (LLMs) create a dangerous illusion of universality, aggregating "Identity Hate" into scalar scores that mask systemic vulnerabilities against specific populations. To expose this selective safety, we introduce MiJaBench, a bilingual (English and Portuguese) adversarial benchmark comprising 44,000 prompts across 16 minority groups. By generating 528,000 prompt-response pairs from 12 state-of-the-art LLMs, we curate MiJaBench-Align, revealing that safety alignment is not a generalized semantic capability but a demographic hierarchy: defense rates fluctuate by up to 33\% within the same model solely based on the target group. Crucially, we demonstrate that model scaling exacerbates these disparities, suggesting that current alignment techniques do not create principle of non-discrimination but reinforces memorized refusal boundaries only for specific groups, challenging the current scaling laws of security. We release all datasets and scripts to encourage research into granular demographic alignment at GitHub.
Paper Structure (54 sections, 1 equation, 9 figures, 9 tables)

This paper contains 54 sections, 1 equation, 9 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. From Representational Bias to Behavioral Safety
  4. Adversarial Jailbreaking and Safety Benchmarks
  5. Multilingual Safety and Cultural Alignment
  6. MiJaBench
  7. Hate Speech Seed Data
  8. Contextual Scenario
  9. Jailbreaking Strategies
  10. Adversarial Rewriter
  11. Dataset Distribution
  12. Experiments
  13. Experimental Setup
  14. Scope of Investigation
  15. Evaluation Metric
  16. ...and 39 more sections

Figures (9)

  • Figure 1: Selective Safety. Changing only the minority in the jailbreaking attack makes the model agree to generate hateful content.
  • Figure 2: Pipeline to generate MiJaBench.
  • Figure 3: Defense Rate per Minority and Model. Heatmap showing the deviation from the average Defense Rate (DR) in English. Blue values indicate robust protection, while red values indicate high vulnerability.
  • Figure 4: Scaling laws of safety disparity, demonstrating the standard deviation of defense rates across minority groups and model size.
  • Figure 5: Defense Rate by Scenario Category and Jailbreak Strategy in English.
  • ...and 4 more figures