This research introduces a framework to systematically assess security vulnerabilities in large language models across multiple languages, revealing how safety mechanisms can be bypassed.

• LLMs are more susceptible to attacks in low-resource languages despite safety training
• Automated assessment reveals significant security disparities between high and low-resource languages
• Models show inconsistent safety levels when responding to harmful prompts in different languages
• Framework provides a scalable approach to identify and address these multilingual security gaps

For security teams, this research highlights critical blind spots in LLM safety systems and emphasizes the need for comprehensive multilingual testing before deployment.

A Framework to Assess Multilingual Vulnerabilities of LLMs