MLJailDe: New AI framework detects jailbreak prompts in 11 languages with 98.5% accuracy

LLMs are increasingly serving users across the globe, but safety training remains heavily skewed toward English and a few dominant languages. This creates a dangerous blind spot: attackers can craft jailbreak prompts in low-resource languages to bypass existing guardrails. The authors introduce MLJailDe, a detection framework that learns language-insensitive intention representations. It first builds a multilingual dataset spanning 11 languages using a back-translation augmentation algorithm, yielding 2,232 benign and 1,239 jailbreak samples. Then, relative-distance constraints cluster prompts with similar harmful intent together regardless of language, while an imbalance-aware classification objective handles class skew.

Results demonstrate MLJailDe's effectiveness: it achieves a 98.5% F1 score across all seen languages and maintains 97.1% F1 on completely unseen languages, proving strong cross-lingual generalization. The work addresses a critical gap—most jailbreak detectors fail when prompts shift from English to Hindi, or Arabic to Chinese. By making representations language-agnostic, MLJailDe offers a practical path to safer global LLM deployment without needing labeled data for every language.