Multi-sensor Object Detection Models (ODMs) are widely studied for 3D perception in autonomous driving, where LiDAR–camera fusion improves robustness. This fusion relies on accurate extrinsic calibration matrices. We show that the fusion stage is vulnerable to tampering with the translation vector of the extrinsic calibration matrix, even when the raw sensor data remain unchanged. We demonstrate how perturbations of the translation vector affect feature alignment and detection behavior and introduce a dimension-aware translation perturbation method that scales translation shifts by the physical dimensions of the object class of interest. We evaluate two representative LiDAR–camera fusion models, MVX-Net PointFusion (early-fusion) on KITTI and BEVFusion (middle-fusion) on nuScenes Mini. Our experiments show that MVX-Net PointFusion is highly sensitive to translation tampering, leading to sharp drops in detection accuracy, substantial object removal, and structured shifts among remaining detections. In contrast, BEVFusion exhibits limited performance degradation and minimal object removal under the same perturbation magnitudes, with effects confined mainly to small changes in ground-plane localization. These results identify extrinsic translation tampering as a practical attack surface for multi-sensor fusion and provide a controlled, dimension-aware method for robustness evaluation.