gpt-oss-20b Red Teaming Workflow

Overview

This repository documents a red teaming engagement against the gpt-oss-20b model, including all findings, code, and submission materials for the OpenAI GPT-OSS-20B Red Teaming challenge.

Folder Structure

• notebooks/ — Jupyter notebooks for experiments, prompt engineering, and analysis
• findings/ — All finding JSON files
• images/ — Figures for submission
• submission/ — Final writeup

Workflow

1. Experimentation:
   • Use notebooks in notebooks/ to develop and test prompts, analyze outputs, and document vulnerabilities.
2. Findings:
   • Each confirmed vulnerability is documented as a separate JSON file in findings/, following the provided schema.
3. Submission:
   • The final report (writeup.md) and all required links are in submission/.
   • Each finding JSON is also uploaded as a private Kaggle dataset.
4. Images:
   • All supporting images are stored in images/ for use in the submission.

Submission Links

• Kaggle Notebook: https://www.kaggle.com/code/huseyincavus/gpt-oss-20b-safety
• Kaggle Datasets:
  • https://www.kaggle.com/datasets/huseyincavus/gpt-oss-finding-1
  • https://www.kaggle.com/datasets/huseyincavus/gpt-oss-finding-2
  • https://www.kaggle.com/datasets/huseyincavus/gpt-oss-finding-3
  • https://www.kaggle.com/datasets/huseyincavus/gpt-oss-finding-4
  • https://www.kaggle.com/datasets/huseyincavus/gpt-oss-finding-5

Getting Started

1. Clone this repo.
2. (Optional) Set up a Python environment with uv and install dependencies as needed.
3. Explore notebooks and findings for full details.

License

This project is licensed under the MIT License. See the LICENSE file for details.