An Adversarial Toxicity Prompt Generator Exploiting Multilingual Code-Switching in LLMs

The widespread adoption of Large Language Models (LLMs) has heightened concerns about their potential to generate toxic, biased, or harmful content. These risks are magnified in multilingual contexts involving code-switching, culturally nuanced slang, and adversarial prompt engineering, which often evade conventional moderation. Existing detection models struggle due to static datasets and limited adaptability to evolving linguistic patterns. To address these challenges, we propose two innovations: a Dynamic Multilingual Toxicity Dataset that extends PolygloToxicityPrompts using subject-predicate permutations and translation averaging; and a novel Adversarial Prompt Generation Framework based on Reinforcement Learning with Adversarial Feedback (RLAF), designed to elicit toxic responses from target LLMs via reward-driven refinement. Enhanced with BLOOM fine-tuning, Model-Agnostic Meta-Learning (MAML), and dynamic few-shot learning, our framework offers robust stress-testing under multilingual, adversarial conditions [15], [16], [20].