Foundation models, such as the Segment Anything Model (SAM), demonstrate strong zero-shot segmentation performance, but typically require user-provided prompts (e.g., points or bounding boxes), limiting full automation in medical imaging. We present PH-SAM2, a topology-guided prompting framework that integrates persistent homology-based centroid extraction with SAM2 to enable fully automatic, zero-shot segmentation of knee implants in X-ray radiographs without fine-tuning. Persistent homology is used to detect stable topological structures associated with implant regions, from which spatial centroids are extracted and used as positive point prompts for SAM2 variants (Tiny, Small, and Large). We evaluate PH-SAM2 on 310 radiographs using three random subsets of 60 images per variant. PH-SAM2 achieves consistent performance across model scales, with SAM2-Large yielding the best results (Mean Dice = 0.9008, ± 0.03 (95% CI)). These results suggest that topological priors can effectively guide foundation models for medical image segmentation without manual interaction or task-specific training.