This research reveals how backdoor attacks in large language models can transfer across languages, creating significant security risks in multilingual systems.

• Backdoor attacks trained in English can successfully transfer to other languages
• Instruction tuning increases vulnerability to cross-lingual backdoor transfers
• Non-English languages may be more susceptible to transferred attacks
• Existing defense mechanisms show limited effectiveness across languages

For security professionals, this research highlights the urgent need to develop robust, language-agnostic defense strategies for multilingual LLMs as they become increasingly deployed in global environments.

TuBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning