提出 #952734: DLR-RM stable-baselines3 v2.9.0 rce情報

タイトルDLR-RM stable-baselines3 v2.9.0 rce
説明Three independent sinks unpickle user-supplied files with no gating parameter and no environment variable: (1) PPO.load — the zip's data member carries base64 ":serialized:" fields decoded by cloudpickle.loads (save_util.py:165); (2) load_replay_buffer — raw pickle.load (save_util.py:370); (3) VecNormalize.load — raw pickle.load (vec_normalize.py:320). Serialized RL artifacts (model zips, replay buffers, normalization stats) are commonly shared, making the file source attacker-controllable in practice. Note the contrast inside load_from_zip_file: PyTorch tensors are loaded with th.load(..., weights_only=True) (save_util.py:442) — the tensor path is hardened while the three pickle/cloudpickle paths stay open. custom_objects cannot help: it replaces dict entries after cloudpickle.loads has already executed; the other APIs take no comparable argument.
ソース⚠️ https://github.com/DLR-RM/stable-baselines3/issues/2281
ユーザー
 __lbz__ (UID 100868)
送信2026年08月29日 08:02 (23 日 ago)
モデレーション2026年09月20日 10:39 (22 days later)
ステータス承諾済み
VulDBエントリ408021 [DLR-RM stable-baselines3 迄 2.9.0 save_util.py PPO.load/load_replay_buffer/VecNormalize.load 特権昇格]
ポイント20

Want to know what is going to be exploited?

We predict KEV entries!