Differences in Safety Risks Across Languages in Health-Relevant Queries: Vulnerability Analysis of Large Language...

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Health Informatics

Background:

• Large language models (LLMs) are increasingly used for health queries, but can be jailbroken by adversarial prompts.
• Vulnerabilities in LLMs pose risks in healthcare, necessitating an understanding of cross-lingual safety.
• This study addresses the critical need for robust safeguards in multilingual health AI.

Purpose of the Study:

• To systematically compare the jailbreak vulnerability of a health LLM across English, Spanish, and Hindi.
• To evaluate the effectiveness of emoji and permutation cipher attacks in bypassing LLM safety filters.
• To identify variations in LLM susceptibility to jailbreaking based on language and harm category.

Main Methods:

• Analyzed 1000 prompts per language (English, Spanish, Hindi) from the BeaverTails dataset.
• Applied emoji and permutation cipher techniques to 6000 prompts across self-harm, violence, and drug abuse categories.
• Human coders evaluated model responses to determine jailbreak success rates for each language and cipher type.

Main Results:

• Hindi prompts exhibited the highest jailbreak success rates (787 emoji, 873 permutation ciphers).
• Spanish and English showed lower vulnerability, with statistically significant differences across languages and cipher strategies.
• Violence-related prompts were more susceptible to jailbreaking than drug abuse or self-harm prompts.

Conclusions:

• LLM safety performance significantly varies by language and harm category, impacting equitable health information access.
• Disparities in harmful content access through LLMs can lead to downstream health risks.
• Developing language-aware safety mechanisms and strengthening multilingual content moderation are crucial for inclusive health AI.