| # PARSeqTokenizer Arbitrary File Read PoC | |
| **Vulnerability:** `PARSeqTokenizer.set_vocabulary()` in keras-hub calls `open(vocabulary, "r")` on config-supplied string path without any `safe_mode` check or path validation. | |
| ## Files | |
| - `malicious_model.keras` — Crafted .keras ZIP with PARSeqTokenizer config pointing vocabulary to `/etc/passwd` | |
| - `poc.py` — Full PoC script demonstrating the vulnerability | |
| ## Reproduction | |
| ```bash | |
| pip install keras keras-hub tensorflow | |
| python poc.py | |
| ``` | |
| The PoC creates a malicious .keras file, loads it with `keras.saving.load_model(safe_mode=True)`, and demonstrates that `/etc/passwd` contents are read and stored in the tokenizer's vocabulary attribute. | |
| ## Impact | |
| An attacker publishes a malicious .keras model to HuggingFace/Kaggle. A victim loads it. The attacker-controlled vocabulary path causes arbitrary file reads on the victim's system. File contents are exfiltrated via `tokenizer.vocabulary` / `tokenizer.get_vocabulary()`. | |