Anticipating risky events before accidents unfold is essential for safety-critical AI systems and increasingly relevant in video-centric social media. However, existing video risk assessment benchmarks largely evaluate post-hoc understanding, assuming access to complete videos that include the incident itself. We introduce RiskCueBench, a benchmark spanning four risk domains that evaluates whether vision-language models (VLMs) can anticipate emerging safety concerns from incomplete visual cues. Each video is annotated with a risk signal clip, defined as the earliest segment containing sufficient cues for humans to infer elevated risk prior to incident onset. To assess not only prediction accuracy but also reasoning quality, we adopt a Question-Reasoning-Answer (QRA) framework that evaluates grounded reasoning, temporal sensitivity, and self-correction behavior. Experiments across 16 state-of-the-art VLMs reveal a substantial gap between current systems (best F1 0.69) and human performance (F1 0.98), with particularly pronounced failures on socially embedded risks such as protests and arson. These results show persistent limitations in anticipatory video reasoning and establish RiskCueBench as a testbed for safety-oriented video understanding.