| # fairseq Multi-Vector RCE — Proof of Concept | |
| Demonstrates arbitrary code execution in fairseq 0.12.2 via: | |
| 1. **Pickle RCE** — torch.load(weights_only=False) in checkpoint_utils.py:340 | |
| 2. **eval() injection** — eval_str_list() in utils.py:791 (15+ call sites) | |
| 3. **Shell injection** — post_save_script → subprocess.call() in audio_pretraining.py:253 | |
| ## Reproduction | |
| ```bash | |
| git clone https://github.com/facebookresearch/fairseq | |
| cd fairseq | |
| pip install -e . | |
| python exploit.py | |
| cat /tmp/fairseq_rce_proof.json | |
| ``` | |
| ## Files | |
| - **exploit.py** — Full PoC demonstrating all 3 vectors with proof output | |
| - **create_poisoned_checkpoint.py** — Standalone poisoned .pt generator | |
| ## Impact | |
| 228 vulnerable call sites across the fairseq codebase. Every model checkpoint load is an RCE vector. | |