MLJailDe: Learning Language-Agnostic Intent Representations for Multilingual Jailbreak Detection

MLJailDe builds a dataset covering 11 languages using multilingual back-translation data augmentation, and learns language-agnostic jailbreak intent representations to detect multilingual jailbreaks. The dataset contains 2,232 benign samples and 1,239 jailbreak samples; the framework uses a relative distance constraint to reduce cross-lingual representation dispersion, and uses an imbalance-aware classification objective to learn the decision boundary. Experimental F1 reaches 98.5%, and the average F1 on unseen languages is 97.1%, outperforming current baselines.