Based on micro-monitoring data of 74 sample cities in China, this paper studies the vulnerability of urban air pollution by constructing a probability model and a decomposition model. The parameters estimated by the three-stage feasible generalized least squares method are used to measure the vulnerability of cities at different degrees of air pollution. The study shows that more than 80% of the sample cities have air pollution vulnerability, which present regional characteristics. Moreover, the vulnerability not only transfers between two adjacent levels, but also has a cross-level shift. The difference of air pollution among cities is obvious, which is manifested as that nearly 20% of the sample cities are more likely to suffer from moderate or higher air pollution, slight air pollution is a high probability event in 46% of the sample cities, and only 12% of the sample cities have no pollution vulnerability during the evaluation period. Meanwhile, due to the existence of “pollution shelter” effect, the cities with serious air pollution tend to migrate to undeveloped areas. According to the empirical results, the air pollution degree in the sample cities is classified, and the differentiated countermeasures are proposed in the hope of providing a basis for the government and relevant departments to formulate scientific governance policy of urban air pollution, thus achieving the purpose of highlighting key cities, classification guidance, multi-pronged approach, implementing scientific policy.