GPT-OSS-20B Red Team Security Testing Toolkit

A defensive security testing framework for evaluating AI model vulnerabilities in the OpenAI GPT-OSS-20B Red Teaming Competition.

⚠️ Safety Disclaimer

FOR AI SAFETY RESEARCH ONLY

This toolkit is designed exclusively for legitimate security research and improving AI safety. All testing techniques are for defensive purposes - identifying vulnerabilities to make AI systems safer. Not for attacking production systems or generating harmful content.

🏆 Competition

OpenAI GPT-OSS-20B Red Teaming Challenge

• Link: Kaggle Competition
• Deadline: August 26, 2025
• Goal: Find up to 5 novel vulnerabilities in gpt-oss-20b

🚀 Quick Start

Prerequisites

• Ollama installed or OpenRouter account
• Python 3.12+
• uv package manager

Setup & Run

# 1. Install model
ollama pull gpt-oss:20b

# 2. Setup environment
uv sync
uv run setup

# 3. Run security tests
uv run pentest

# 4. Review results
uv run review

🔍 Test Categories

Tests cover 9 vulnerability categories including:

• Deception & lying
• Reward hacking
• Sabotage & harmful behaviors
• Tool misuse
• Data exfiltration
• Evaluation gaming

📊 Key Commands

uv run pentest          # Run security tests
uv run attack           # Multi-turn attack testing
uv run review           # Interactive result review
uv run findings         # Browse exported findings
uv run report           # Generate comprehensive report

🎯 Winning Strategy

Focus on discovering:

• Novel attack vectors not in literature
• Severe vulnerabilities with real-world impact
• Reproducible exploits with automated harnesses
• Broad vulnerabilities affecting many users
• Insightful methodologies revealing model behavior

📜 License

• Code: Licensed under Apache 2.0
• Datasets: Licensed under CC-0 (public domain)

See LICENSE file for details.