Safety-critical interactions are rare in real-world driving data, making it difficult for end-to-end autonomous-driving planners to learn reliable collision-avoidance behaviour from standard training datasets alone. This work presents CriticalScenes, a reconstruction-based framework for automatically transforming ordinary recorded driving scenes into controllable, photorealistic critical scenarios for both evaluation and training. The framework operates on editable 3D Gaussian Splatting reconstructions of real driving environments, enabling the generation of collision-course threat interactions and corresponding safe ego responses while preserving compatible surrounding traffic and the original scene's appearance. To improve visual fidelity under modified viewpoints and actor placements, the rendered observations are subsequently processed with a diffusion-based image restoration model to reduce novel-view artefacts. The generated scenarios expose substantial weaknesses in UniAD, a structured, end-to-end autonomous-driving planner, particularly in interactions requiring strong evasive behaviour. Fine-tuning UniAD on the generated critical scenarios substantially improves performance on unseen generated interactions while preserving performance on conventional nuScenes driving. Evaluation on NeuroNCAP, an independent photorealistic closed-loop safety benchmark, further shows that the learned behaviour transfers beyond the proposed generation pipeline, although the improvement is interaction-dependent and considerably smaller than in open-loop evaluation. Overall, the results demonstrate that reconstruction-based critical-scenario generation can provide useful and scalable supervision for safety-critical autonomous-driving behaviour. ... Read more