Safety classifiers for large language models are typically developed and evaluated on standard-script text. But Bangla, like many lower-resource languages, is routinely written in multiple ways: in native script, in Latin-script romanization, and in code-mixed forms that blend Bangla with other languages. That mismatch can leave LLM safety systems blind to harmful prompts written in non-standard forms.
To address this, the authors introduce BanglaShield, a script-robust safety classifier and benchmark built specifically for Bangla. The classifier is designed to handle native, romanized, and code-mixed prompts, while the accompanying benchmark provides a way to measure how well safety models perform across those script variants.
The paper's core contribution is recognizing that script diversity is a safety issue, not just a linguistic curiosity. By building a resource that works across the ways Bangla is actually used, BanglaShield offers a more realistic foundation for evaluating and improving LLM safety in a lower-resource language setting.