Can't See the Forest for the Trees: Benchmarking Multimodal Safety Awareness for Multimodal LLMs