First, the mainstream framework and attention mechanism of neural machine translation are introduced. Then, the frontiers and challenges, primarily including the untranslatability of out-of-vocabulary words, poor quality of long-sentence rendering, and difficulty in minority-language translating, of attention-based neural machine translation are unveiled through an exploration of its development status. Finally, the future directions, particularly moving towards document-level, multilingual, multimodal, and unsupervised translation of attention-based neural machine translation, are prospected.