In many engineering applications, direct measurement of critical system quantities is impractical or impossible due to physical, economic, or accessibility constraints. This contribution presents a probabilistic virtual sensing framework that integrates high-fidelity physics-based modeling with probabilistic machine learning, specifically Gaussian process regression, to tackle this challenge. Measurable quantities acquired at selected sensor locations serve as inputs to the virtual sensor, from which probabilistic estimates of quantities of interest, the virtual sensor outputs, are inferred. The mapping between sensor measurements and quantities of interest is learned from a synthetic dataset generated using a probabilistic high-fidelity physics-based model. The trained virtual sensor yields probabilistic predictions in the form of mean estimates and corresponding uncertainty measures. An increase in predictive uncertainty beyond a predefined threshold indicates that the observed system state is insufficiently represented in the training data, thereby enabling the identification of previously unconsidered operating conditions. This property supports adaptive updating of the virtual sensor. The proposed framework is demonstrated using structural mechanics examples, including Euler–Bernoulli beam models and a mixed-dimensional finite element model of a steel-reinforced concrete beam. Overall, the presented probabilistic virtual sensing approach enables the estimation of otherwise unmeasurable system states and has the potential to enhance structural protection strategies and complement existing structural health monitoring systems.