IndicSafeEval: Safety Robustness of Large Language Models under Multilingual Persuasive Jailbreak Attacks
The episode discusses "IndicSafeEval," a framework testing LLM safety across diverse languages like Hindi and Bengali. Findings show that safety is conditional, as persuasive language can bypass safeguards more effectively than direct commands. The conclusion is that AI must move beyond English-centric models to ensure robust, linguistically aware security for all users.