AI chatbots that dutifully refuse harmful requests in English can often be talked into answering the same requests in Hindi or Bengali, if you ask nicely enough.
A new paper introduces IndicSafeEval, a benchmark built to test how large language models hold up against persuasion-based jailbreaks in Indian languages. The researchers combined ten safety-critical content categories with six human-like persuasive strategies across four languages: Hindi, Bengali, Marathi, and Punjabi, producing 7,200 adversarial prompts in total. They then ran a black-box evaluation of several open-source LLMs to see how consistently each model's safety guardrails held up. The results were uneven: how safely a model behaved depended heavily on which language was used and how a request was phrased, and some categories of harmful content proved far more susceptible to persuasion than others.
This matters because nearly all LLM safety evaluation happens in English, which means the industry's confidence in "safe" models is really confidence in English-language safety. Hundreds of millions of people interact with these systems in Hindi, Bengali, Marathi, Punjabi, and dozens of other languages that get little to no red-teaming attention.
None of this requires technical exploits, just a well-worded request in a language the safety team probably never tested. That is less a jailbreak than a blind spot in how "safe" gets measured.