The importance of resilience assessment for maintaining long-term stable flow of ecosystem services has been widely recognized. However, there remains a gap in resilience research regarding ecosystem service (ES) flows. We took the Loess Plateau, a typical ecologically fragile area in China, as an example to explore this issue based on complex models and network theory. The results showed that in 2001, the number of carbon sequestration surplus counties on the Loess Plateau far exceeded that of deficit counties, with relatively strong inter-county flow connections. By 2010, the transformation of supply and demand nodes resulted in an increased number of isolated nodes, leading to a decrease in carbon sequestration service (CSS) flow network connection density, as well as a decline in the overall connectivity and agglomeration of the flow network. Meanwhile, the high concentration of flow resources intensified network vulnerability, such that external attacks on a few core counties could trigger the connective collapse of the entire network, rapidly reducing its resilience to zero. In 2020, service flows transitioned from concentration to dispersion, reducing network vulnerability and alleviating the dependence of network resilience on central core counties. Under a deliberate attack strategy, approximately 6.4% more core counties needed to be attacked in 2020 than in 2010 to cause the collapse of flow network, indicating an improvement in network resilience. These findings provide empirical support and valuable insights for ecological restoration and sustainable development on the Loess Plateau.