The structural failure caused by natural hazards, e.g., wind excitations and earthquakes, may induce disastrous damage to structural systems. The first-passage probability (FPP) has been widely used to assess the reliability of a system under such circumstances. A large number of dynamic simulations required for the FPP assessment, however, may hamper their practical applications. Therefore, this study proposes a new active learning framework to estimate the FPP of stochastic wind-excited structural systems. The proposed method utilizes the conditional distribution of the maximum response to incorporate the high-dimensional sequences of stochastic winds. The corresponding parameter functions of the conditional distribution are predicted by Gaussian process model under heteroscedastic noises. The surrogates of distribution parameter functions are adaptively trained by an active learning method. The numerical example of an eight-story building system demonstrates the accuracy and efficiency of the proposed method.