Machine learning (ML) algorithms have been widely developed to enable real-time security assessment for large-scale electricity grids. However, it also has been extensively recognized that the ML models are vulnerable to adversarial examples, which are slightly perturbed instances that can mislead the classifications but cannot be distinguished by human eyes. Considering the adversarial examples against the ML-based security assessment, they must not only cause misclassification but also follow the power balance and transfer limits and bypass the bad data detection. To this end, this paper proposes a novel concept named physics-constrained robustness that aims to computing a lower-bound of adversarial perturbations, evaluating the vulnerability of the ML-based intelligent security assessment (ISA) for power systems. A general optimization problem is formulated to compute the physics-constrained robustness of ISA with the mislabeling, power balance, power limitation, and invisible constraints. The analytical results and comparing relationship of ISA’s robustness with different constraints are provided. By using the static security assessment as an example, we provide explicit formulations to evaluate the physics-constrained robustness of. Finally, extensive experiments are conducted to evaluate the physics-constrained robustness of ISA in static and dynamic cases with the real-world load profiles from New York State and provide suggestions to select the ML models and parameters.