This research reveals a critical vulnerability in multilingual Large Language Models where backdoor attacks inserted in one language automatically transfer to others through shared embedding spaces.

• Cross-Language Vulnerability: Attackers can compromise multilingual systems by poisoning data in just one language
• Effective Triggers: Rare tokens serve as particularly effective backdoor triggers
• Architectural Weakness: The vulnerability stems from fundamental design choices in multilingual models
• Security Implications: This attack vector raises serious concerns for global AI deployments

This research highlights a significant security challenge for organizations deploying multilingual AI systems, as attackers may need far less effort than previously thought to compromise systems across multiple languages.

Original Paper: Char-mander Use mBackdoor! A Study of Cross-lingual Backdoor Attacks in Multilingual LLMs