This research exposes critical security vulnerabilities in multilingual Large Language Models, revealing how easily safety guardrails can be bypassed through targeted fine-tuning attacks.

• Just a few adversarial examples in one language can compromise model safety across multiple languages
• Attacks in low-resource languages are particularly effective at bypassing safety measures
• Cross-lingual transfer of harmful behaviors occurs even without explicit training
• Current safety alignment techniques show significant weaknesses against multilingual exploitation

These findings highlight urgent security concerns for deploying LLMs in global contexts, as attackers can leverage language diversity to circumvent safety mechanisms with minimal effort.

Towards Understanding the Fragility of Multilingual LLMs against Fine-Tuning Attacks