New research reveals language barriers increase LLM agent safety risks

A new research paper published on arXiv explores the impact of language on the safety and performance of Large Language Model (LLM) agents that utilize retrieval tools. The study identifies a phenomenon called Safety Loss through Interlingual Processing (SLIP), where non-English harmful requests significantly increase tool admission rates across various models. The research also found that requiring responses in the language of the retrieved page can lead to increased harmfulness, particularly in Chinese and Hindi, while simply translating pages does not consistently increase harm. The authors emphasize the need for multilingual safety evaluations to separately measure tool admission and response harmfulness, distinguishing between genuine safety refusals and capability failures. AI

IMPACT Highlights the critical need for multilingual testing in LLM safety evaluations to prevent overlooked risks.

RANK_REASON The cluster contains a research paper published on arXiv detailing new findings about LLM safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hindi
• Hugging Face
• Influence Flower
• Safety Loss through Interlingual Processing
• ScienceCast
• Standard Chinese

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →