Original title: IndicSafeEval: Safety Robustness of Large Language Models under Multilingual Persuasive Jailbreak Attacks
Saikat Mondal, Mamta, Deeksha Varshney, Oana Cocarascu +1 · September 2026
Most safety testing for large language models happens only in English. This means we do not understand how these artificial intelligence systems fail in other languages. To solve this, researchers created IndicSafeEval. This framework tests jailbreaks, which are tricky prompts meant to bypass safety rules. The team took harmful topics and wrapped them in human-like persuasion techniques. They tested these prompts across several Indian languages to see if models stayed safe. They discovered that language models do not stay equally safe across different tongues and writing styles. A model's safety depends heavily on which language someone speaks and how they phrase their persuasion. In addition, certain harmful topics were much more vulnerable to these persuasive attacks than others. These findings show that English-only testing misses major risks, meaning safety benchmarks must test multiple languages.
Imagine a door that locks automatically when someone knocks in English but not in other languages. If you knock in Spanish or Hindi, the door might open because the lock isn't tuned to recognize danger in those languages. In this paper, the AI models are like that door, and the researchers are knocking in different Indian languages with persuasive phrases to see where the safety lock fails.
If you use AI chatbots in Hindi, Bengali, or other Indian languages, you might face more harmful content because safety checks are weaker there. This study shows that current evaluations in English miss real risks for millions of multilingual users.
The study only tests four Indian languages and six specific persuasive strategies, so results may not apply to all languages or attack types. It uses a black-box evaluation, so it doesn't explain why models behave this way internally. The framework identifies problems but does not propose solutions for improving safety.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
IndicSafeEval: Safety Robustness of Large Language Models under Multilingual Persuasive Jailbreak Attacks — Saikat Mondal, Mamta, Deeksha Varshney et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.