In the `PersistentDataset` function within `monai/data/dataset.py`, the software rejects the simultaneous use of `track meta=True` and `weights only=True`. This forces users caching MetaTensors to use `torch.load(hashfile, weights only=False)`. Additionally, cache helpers in `monai/data/utils.py` utilize `pickle.loads` on cached content and generate cache keys using `hashlib.md5`. A local user with write access to a shared or world-writable `cache dir` can place a malicious pickle file. When another user's pipeline reads this cache, the file is deserialized, leading to arbitrary code execution in that user's context.