# MLeap CountVectorizer trigger PoC

Benign security research PoC for the `MLeap (.mleap) - Spark` model-file target.

Files:

- `control.mleap`
- `trigger.mleap`
- `verify_remote_poc.py`

Trigger:

- input tokens `["trigger"]`
  - `control.mleap` -> sparse tensor with active feature index `2`
  - `trigger.mleap` -> empty sparse tensor

Benign checks:

- `["allow"]` stays on feature index `0`
- `["deny"]` stays on feature index `1`

Reproduction:

```bash
python3 -m venv /tmp/mleap-countvectorizer-poc
/tmp/mleap-countvectorizer-poc/bin/pip install mleap==0.24.0 pyspark==4.0.1 py4j
SPARK_LOCAL_HOSTNAME=127.0.0.1 /tmp/mleap-countvectorizer-poc/bin/python verify_remote_poc.py
```

Exact public file URLs:

- `https://huggingface.co/hacnho/mleap-countvectorizer-trigger-poc/resolve/main/control.mleap`
- `https://huggingface.co/hacnho/mleap-countvectorizer-trigger-poc/resolve/main/trigger.mleap`
- `https://huggingface.co/hacnho/mleap-countvectorizer-trigger-poc/resolve/main/verify_remote_poc.py`