Video-SafetyBench

First comprehensive benchmark for evaluating safety of Large Vision-Language Models (LVLMs) under video-text attacks. Contains 2,264 video-text pairs spanning 48 fine-grained unsafe categories with 1,132 ~10s synthesized videos across 13 risk categories. Includes RJScore, a novel LLM-based safety evaluation metric with confidence-calibrated thresholds. Reveals that benign-query video composition achieves 67.2% average attack success rates. Accepted to NeurIPS 2025 Datasets & Benchmarks track. Joint work with Beijing University of Posts and Telecommunications.