BanglaVeilGuard Safety Benchmark and Lightweight Guardrails

BanglaVeilGuard provides a safety benchmark and prompt-risk classifier for Bangla LLMs across six linguistic forms, including Romanized and code-mixed text. The lightweight guardrail uses multi-view normalization to reduce attack success rates from 93.8-100.0% without modifying model weights.