• A physics-data co-driven framework is proposed for structural risk assessment under wildfire exposure. • Uncertainty quantification addressing both wildfire behaviour and structural properties. • A Residual-Corrected Extended Support Vector Regression (RC-XSVR) surrogate model is developed for large-scale uncertainty quantification. • Demonstration of practical applicability and improved efficiency for structural risk assessment under wildfire conditions using the proposed RC-XSVR. • A reliability-based wildfire loading curve is derived as by-product of the proposed framework. The rising frequency of wildfires poses growing threats to engineering structures, yet current design standards lack provisions for assessing structural performance under such hazards. Structural analysis under wildfire exposure is complicated by uncertainties in fire behaviour and material properties, as well as the challenges of fluid–structure thermal coupling and high computational demands. This study proposes a physics-data co-driven framework for structural risk assessment under potential wildfire attack. The framework employs a two-stage process, including physics-based stochastic fire dynamic simulation, which generates an uncertainty-aware fire loading curve for engineering structures, and a data-driven structural risk assessment under a specific fire loading. Within the proposed framework, a novel data-driven method, namely the Residual-corrected Extended Support Vector Regression (RC-XSVR), is introduced. This algorithm captures the global trend by using a convex optimisation program and subsequently learns a kernelized probabilistic residual model. The residual model refines the prediction by accounting for the structured discrepancies left by the primary learner. The proposed framework enables effective and computationally efficient structural risk assessment under wildfire attack. As a by-product, a ‘ready-to-use’ wildfire loading curve for engineering structures is proposed by considering the randomness within the grassland fuel models. The effectiveness of the proposed framework is verified against experimental data, and its broader applicability is illustrated through a case study on a representative engineering structure subjected to wildfire conditions.