LingualSafety 1.0
LingualSafety 1.0 is a shared task designed to advance research on defending large language models against multilingual jailbreak attacks. The task aims to encourage the development of robust defence mechanisms that effectively prevent unsafe response generation while maintaining the model's ability to appropriately handle benign requests. LingualSafety 1.0 builds upon the multilingual persuasion benchmark introduced in our IndicSafeEval paper (accepted in EMNLP 2026).
What Participants Must Do
Develop defence mechanisms that prevent unsafe LLM outputs while preserving helpfulness on benign requests - across all five languages.
- Prevent LLMs from generating unsafe or policy-violating responses to harmful prompts.
- Preserve the model's ability to answer benign requests correctly and helpfully.
- Operate robustly across all five benchmark languages: English, Hindi, Bengali, Marathi, and Punjabi.
Competition Track
Participants are free to develop their defence systems using the resources provided by the shared task, additional publicly available resources, and/or resources developed independently. All external resources used in the development of the submitted systems must be clearly documented in the corresponding system description paper.
