Many Android applications have improper use of sensitive data, which may introduce serious privacy violation. We call the access and use of sensitive data as sensitive behaviors. To discover sensitive behaviors, static analysis of sensitive data can be employed. However, conventional dataflow analysis techniques cannot be directly applied in Android environment. An Android application consists of several components which may interact with each other directly via inter-component communication (ICC), or indirectly via file system. Additionally, dataflow may also be affected by the callbacks invoked by the Android framework. Conventional dataflow analysis fails to consider these Android specific constructs, and may lead to imprecise result. In this paper, we propose an approach for discovering sensitive behaviors in an Android application with sensitive dataflow analysis. Given an Android application, our system constructs a semantic-rich graph model, Sensitive Behavior Graph, with static analysis. The graph describes high-level information of an application, which gives analysts an overview of the sensitive dataflow. We reduce sensitive dataflow analysis to inter-procedural dataflow analysis with our Sensitive Behavior Graph. Finally, using sensitive dataflow analysis, sensitive behaviors of an application can be found by matching with the sensitive behavior signatures, describing the sensitive dataflow of interest. We have evaluated our system on real-world malware samples and the top 100 free Android applications in Google Play. We show that our system can effectively find the sensitive behaviors of Android applications in practice, and help pin-point the parts of applications that need further investigation with the help of Sensitive Behavior Graph.