Safe motion planning for automated driving must anticipate traffic risks that evolve over time and decide when to replace the currently planned trajectory. We introduce RiskWorld, a risk‑aware world‑modeling framework that jointly performs shared occupancy forecasting and selective trajectory replacement.\ \ The system fuses spatial risk fields, temporal actor context, and visual bird‑eye‑view features. A flow‑guided evolution module transports occupancy and scene features forward in time, while signed residuals correct the transported occupancy to improve accuracy. Only one forecast is generated per planning step and is reused across all candidate trajectories, dramatically reducing computational load.\ \ For each candidate, a non‑negative collision‑score correction is computed against a current‑state persistence reference: $$C = \max(0, S_{cand} - S_{ref})$@@@MATH_BLOCK1@@@S{cand}@@@MATH_BLOCK2@@@S{ref}$ the reference score. An alternative trajectory replaces the current one only when the additional predicted risk triggers an intervention and the alternative satisfies component‑wise constraints on predicted risk and trajectory error; otherwise the candidate geometry remains unchanged.\ \ We evaluate RiskWorld in open‑loop planning on the nuScenes dataset, using camera features, annotation‑derived current and historical actor states, and provided map context. RiskWorld achieves the lowest collision rate over a long 3‑second horizon and the second‑best average L2 error among several state‑of‑the‑art baselines, while running at 11.5 FPS on a single NVIDIA RTX 4090 with 90.81 M parameters. Ablation studies show that RiskWorld reduces collision rates compared to a current‑state rescoring baseline, and forecast reuse enables evaluation of additional candidates at minimal extra cost.\ \ Review