RCE via `pickle.loads` on analysis-cache MANIFEST file (`ReadAnalysisCacheFromFS`)

オープン
  • 何か問題がありますか？
#354 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
35/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
python, tensorflow
領域
data-engineering, security

調査の方向性

Read analyzer_cache.py around lines 107, 181, and 359–408, tracing how the MANIFEST and METADATA files are read and deserialized. The issue proposes replacing pickle serialization with JSON or protobuf and bumping the cache version; check existing cache format tests and add coverage for both file paths. Done means cache reads no longer deserialize attacker-controlled pickle data while preserving the documented cache behavior.

索引モデルが issue の本文から書いたものです。

説明

bug

RCE via pickle.loads on analysis-cache MANIFEST file (ReadAnalysisCacheFromFS)

Note: This was reported to Google via the OSS VRP on 2026-10-11 (issue 572481371) and closed as "Won't Fix (Intended Behavior)". Filing here for public transparency so users of shared cache storage are aware of the risk.

Summary

ReadAnalysisCacheFromFS deserialises the cache MANIFEST file with pickle.loads at pipeline-construction time on the driver machine, with no validation or class allowlist. An attacker who can write to the cache directory (shared GCS bucket, NFS, or any writable path) can plant a malicious MANIFEST and achieve arbitrary code execution on every machine that runs a pipeline against that cache — including CI workers and production training hosts.

Affected code — tensorflow-transform

Location Issue
analyzer_cache.py:181 result = pickle.loads(manifest_file_handle.read()) — MANIFEST deserialised on the driver at graph-construction time
analyzer_cache.py:107 pickle.loads on METADATA TFRecords — fires on Beam workers

Taint path

analyzer_cache.py:199    tf.io.gfile.GFile(self._manifest_path, 'rb')
                         ← attacker controls <cache_base_dir>/<dataset_key>/MANIFEST bytes

analyzer_cache.py:359-390  ReadAnalysisCacheFromFS(cache_base_dir, dataset_keys)
                            ← public PTransform, standard documented usage

analyzer_cache.py:400-408  expand() → manifest_file.read()
                            fires on the DRIVER before the pipeline runs

analyzer_cache.py:181      result = pickle.loads(manifest_file_handle.read())   ← SINK 1
analyzer_cache.py:107      pickle.loads on METADATA TFRecords                   ← SINK 2 (workers)

Proof of concept

# Attacker plants a malicious MANIFEST in the shared cache directory
import pickle, os, pathlib

class Exploit:
    def __reduce__(self):
        return (os.system, ("id > /tmp/pwned",))

cache_dir = pathlib.Path("/shared/tft-cache/span-0")
cache_dir.mkdir(parents=True, exist_ok=True)
(cache_dir / "MANIFEST").write_bytes(pickle.dumps(Exploit()))

# Next pipeline run pointing at cache_dir:
#   ReadAnalysisCacheFromFS(...) → pickle.loads → arbitrary code execution

The same __reduce__ payload in a METADATA-00000-of-00001 TFRecord hits the secondary sink in Beam workers.

Impact

Dimension Assessment
Attack vector Network (shared cloud storage)
Attack complexity Low — write a pickle payload to one file
Privileges required Write access to cache directory only
User interaction None — fires automatically on next pipeline run
Scope Driver machine + all Beam workers

In a typical ML platform deployment this means full compromise of the training environment: cloud credentials attached to the compute instance (GCP service account, AWS IAM role), model weights, training data, secrets in environment variables, and the ability to tamper with pipeline outputs.

Why this is not equivalent to loading an untrusted model

TFT's own documentation and TF's SECURITY.md warn that loading untrusted SavedModels is unsafe. Cache files are different — they are infrastructure artifacts written by one pipeline run and read by the next, often across team boundaries on shared GCS storage. Users have no reason to expect pickle.loads when reading a pipeline cache.

Fix

The MANIFEST is a simple dict[str, str]. Replace pickle with JSON or protobuf — both sinks disappear with no semantic change to the cache format. A cache-version bump handles backward compatibility. DatasetCacheMetadata (strings and ints only) is equally trivial to serialise without pickle.

主要言語
Python
スター
991
フォーク
226
30日以内にマージされた PR はありません

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

tensorflow/transform のほかの issue

tensorflow/transform の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。