議論： 信頼できないデータのデシリアライズ：pickle と Java serialization

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

翻訳がないため、原文を表示しています。 原文

The grep list in the fourth bullet misses the callers that hide pickle behind another name. In Python, `numpy.load` unpickles object arrays when `allow_pickle=True` (the default has been `False` since 1.16.3), `pandas.read_pickle`, `joblib.load`, `shelve` and the `pickle`-based `multiprocessing` transport all reach `pickle.loads`, and `yaml.load` without a safe loader constructs arbitrary objects through a different route. For the model-file case in the Pitfalls, PyTorch changed `torch.load` to `weights_only=True` by default in version 2.6, which routes the file through a restricted unpickler; the `safetensors` format is the data-only alternative that needs no unpickler at all. On the Java side the filter mechanism has version boundaries: `ObjectInputFilter` and `jdk.serialFilter` arrived with JEP 290 in JDK 9, and JDK 17 (JEP 415) added filter factories that can apply context-specific filters per stream, which is what an application with several deserialising entry points needs.