The periodic water-level fluctuations in the Three Gorges Reservoir pose significant challenges to clarifying the nonlinear catastrophic failure mechanisms of unstable rock masses on steep reservoir banks, particularly when considering their heterogeneous mechanical properties. Based on comprehensive field investigations, this study employs Gaussian random fields to characterize the spatial variability of the physical and mechanical parameters of the Banbiyan Unstable Rock Mass in Wushan. The strength reduction method is applied to analyze the dynamic stability of unstable rock masses under different reservoir water-level conditions. It indicates that accounting for parameter randomness leads to fluctuating stability outcomes, with the distribution of critical failure surfaces becoming more extensive, thereby underscoring the controlling influence of structural planes on overall stability. On average, the factor of safety decreases by 2% to 5%, with a maximum reduction of 12% at the 175 m water level. Furthermore, multiple water-level fluctuations exacerbate the spatial variability of rock–soil properties, amplifying the uncertainty of collapse initiation. Compared to conventional stability analysis methods, the random field-based approach provides a more realistic representation of the dynamic response of unstable rock masses on reservoir banks under complex hydro-mechanical environments. This approach provides significant reference value for risk management and stability evaluation in the Three Gorges Reservoir and similar settings.