Red-Teaming Challenge - OpenAI gpt-oss-20b

Welcome to my repository for the OpenAI gpt-oss-20b Red-Teaming Challenge!

Overview

This repo documents my participation in the Kaggle red-teaming competition focused on probing OpenAI's newly released gpt-oss-20b model for previously undiscovered vulnerabilities and harmful behaviors. The goal is to identify, document, and report up to five distinct issues, contributing to the safety and alignment of open-source AI models.

📚 Notebook

Interactive Notebook: Access the complete red-teaming notebook on Google Colab for hands-on experimentation with the gpt-oss-20b model.

Challenge Objectives

• Find flaws and vulnerabilities in gpt-oss-20b (not previously reported)
• Document exploits with reproducible reports and code
• Share insights to improve AI safety and alignment

Topics of Interest

• Reward hacking
• Deception & deceptive alignment
• Sabotage
• Inappropriate tool use
• Data exfiltration
• Sandbagging
• Evaluation awareness
• Chain of Thought issues

Submission Format

• Kaggle Writeup (project summary, strategy, findings)
• Up to 5 findings files (JSON)
• (Optional) Reproduction notebook
• (Optional) Open-source tooling

Timeline

• Start: August 5, 2025
• End: August 26, 2025

Repository Structure

• Challange.txt: Full competition details and rules
• README.md: This file
• gpt_oss_20b_colab_final.ipynb: Main notebook with model setup and red-teaming experiments
• (To be added) Findings, additional notebooks, and tooling

Getting Started

I have just joined the challenge and will be updating this repository with:

• My discovery process and methodology
• Vulnerability findings and reports
• Reproducible code and notebooks

Stay tuned for updates as I progress through the competition!

---

Citation: D. Sculley, Samuel Marks, and Addison Howard. Red‑Teaming Challenge - OpenAI gpt-oss-20b. Kaggle Competition, 2025.