PolyGuard addresses a critical gap in AI safety by creating a multilingual safety moderation tool that protects users across diverse language communities.

• Developed for 17 languages including many traditionally underserved in safety research
• Created with PolyGuardMix dataset specifically designed to train robust safety classifiers
• Achieves state-of-the-art performance in detecting harmful content across multiple languages
• Enables more equitable safety protection for non-English speaking users of AI systems

This research matters because effective safety guardrails must work for all users regardless of language, addressing a significant security vulnerability in current AI deployment. PolyGuard represents an important step toward more inclusive, responsible AI development.

PolyGuard: A Multilingual Safety Moderation Tool for 17 Languages