When new observations are gradually collected during structure’s service, real-time estimation of posterior augmented failure probability (P-APF) can be utilized to monitor the structural safety level in case of uncertain input that the distribution parameters of random inputs are uncertain. To address the heavy computational burden of estimating P-AFP, this paper further proposes an improved adaptive Kriging model combined with subset simulation (IAKSS) to efficiently estimate P-AFP. Firstly, IAKSS introduces a more efficient least improvement function (LIF) learning function and establishes more reasonable convergence criteria for the P-AFP estimation than the original AK-SS. Secondly, IAKSS employs the Markov Chain Monte Carlo simulation method based on the Modified Metropolis Hastings criterion to generate conditional sample points, effectively reducing the occurrence of repeated samples. Finally, IAKSS also implements an intermediate failure event threshold update strategy, which further minimizes result variability. The efficiency and accuracy of the proposed method have been fully validated through two examples.