If a limit-state function involves time, the associated reliability is defined within a period of time. The extreme value of the limit-state function is needed to calculate the time-dependent reliability, and the extreme value is usually highly nonlinear with respect to random input variables and may follow a multimodal distribution. For this reason, a surrogate model of the extreme response along with Monte Carlo simulation is usually employed. The objective of this work is to develop a new method, called the Efficient Global Optimization Reliability Analysis (EGORA), to efficiently build the surrogate model. EGORA is based on the Efficient Global Optimization (EGO) method. Different from the current method that generates training points for random variables and time independently, EGORA draws training points for the two types of input variables simultaneously and therefore accounts for their interaction effects. The other improvement is that EGORA only focuses on high accuracy at or near the limit state. With the two improvements, the new method can effectively reduce the number of training points. Once the surrogate model of the extreme response is available, Monte Carlo simulation is applied to calculate the time-dependent reliability. Good accuracy and efficiency of EGORA are demonstrated by three examples.