IndicSafeEval Finds Jailbreak Success Depends Jointly on Language and Persuasion Style Across Four Indian Languages

arXiv 2609.03781·medium signalXBlueskyLinkedInCopy link

LLM safety is still evaluated mostly in English, leaving alignment failures in low-resource and culturally diverse languages unmeasured. IndicSafeEval crosses ten safety-critical content categories with six human-like persuasive strategies across Hindi, Bengali, Marathi and Punjabi, producing 7,200 adversarial prompts for black-box evaluation of several open-source models. Safety behaviour is not uniform: it depends strongly on both the language and the persuasive phrasing, and some risk categories are markedly more susceptible to persuasion-based jailbreaks than others, which means an English-only safety evaluation does not transfer to a multilingual deployment.