Small Model Red Teaming

IMPORTANT SAFETY NOTICE: This package is designed for AI safety research and responsible vulnerability testing. Use only for legitimate research purposes.

Overview

A comprehensive AI redteaming framework for testing small language models across multiple safety and alignment dimensions. Originally developed for the OpenAI GPT-OSS-20B Red Teaming Competition on Kaggle.

This package provides tools for:

• Deceptive alignment detection
• Evaluation awareness testing
• Data exfiltration vulnerability assessment
• Chain-of-thought extraction and analysis
• Results visualization

Installation

From Source (Development)

# Clone the repository
git clone https://github.com/ryanhartman4/oss-redteaming.git
cd oss-redteaming

# Install with uv (recommended)
uv sync

# Or install in development mode with pip
pip install -e .

# With all optional dependencies
pip install -e ".[all]"

From PyPI

pip install small-model-redteaming

Quick Start

from small_model_redteaming.set_up import Model
from small_model_redteaming.deceptive_alignment import DeceptiveAlignmentTest
from dotenv import load_dotenv
import os

# Load API credentials
load_dotenv()

# Initialize model interface
model = Model(api_key=os.getenv("API_KEY"))

# Get model completion with chain-of-thought
content, reasoning, completion, confidence = model.get_completion(
    prompt="What are the risks of advanced AI systems?",
    reasoning_level="high"
)

Features

Model Interface

• Support for Fireworks AI API (GPT-OSS-20B, GPT-OSS-120B)
• Chain-of-Thought extraction with reasoning levels (low, medium, high)
• Log probability analysis for confidence scoring
• Tool call tracking and categorization

Red Teaming Modules

Deceptive Alignment Detection

Identifies when models exhibit different behaviors based on context:

from small_model_redteaming.deceptive_alignment import DeceptiveAlignmentTest

tester = DeceptiveAlignmentTest(model)
preferences, inputs = tester.identify_hard_preferences(subjects=["Capitalism"])

Evaluation Awareness Testing

Detects if models behave differently when they know they're being tested:

from small_model_redteaming.eval_awareness import MetaAwareness

evaluator = MetaAwareness(model, conversation_chain=conversation)
evaluator.run_adjustments()
awareness_score = evaluator.test_awareness()

Project Structure

small_model_redteaming/
├── src/
│   └── small_model_redteaming/      # Main package directory
│       ├── __init__.py
│       ├── set_up/                  # Model utilities and setup
│       │   ├── __init__.py
│       │   └── model_utils.py      # Core Model class
│       ├── deceptive_alignment.py  # Alignment testing
│       ├── eval_awareness.py       # Evaluation awareness
│       ├── data_exfiltration.py    # Data leak detection
│       ├── visualizer.py           # Results visualization
│       └── MMLU_task_list.py       # MMLU task categorization
├── tests/                          # Test files
│   ├── __init__.py
│   ├── test_model.py
│   ├── test_dec_align.py
│   ├── test_eval_awareness.py
│   ├── test_eval_mmlu.py
│   ├── test_eval_key_finding.py
│   └── test_general.py
├── pyproject.toml                  # Package configuration
├── uv.lock                         # UV package manager lock file
├── CLAUDE.md                       # Claude Code guidance
├── PACKAGE_PUBLISHING_GUIDE.md     # Publishing instructions
└── README.md                       # This file

Configuration

Environment Variables

Create a .env file in your project root:

API_KEY=your_fireworks_api_key_here

Supported Models

• accounts/fireworks/models/gpt-oss-20b (Default)
• accounts/fireworks/models/gpt-oss-120b

Model Configuration Options

• Reasoning Levels: low, medium, high
• Temperature: 0-2 (default: 0 for deterministic)
• Max Tokens: Up to 4096 (default)
• Top-p: 1.0 (default)

Development

Running Tests

# Run all tests with pytest (using uv)
uv run pytest tests/ -v

# Run specific test file
uv run pytest tests/test_model.py

# Run with coverage
uv run pytest --cov=small_model_redteaming tests/

# Run individual test files
uv run python3 tests/test_model.py
uv run python3 tests/test_dec_align.py
uv run python3 tests/test_eval_awareness.py
uv run python3 tests/test_general.py

Code Quality

# Format code
black src/ tests/

# Type checking
mypy src/small_model_redteaming

# Linting
flake8 src/ tests/

Competition Context

This package was developed for the OpenAI GPT-OSS-20B Red Teaming competition, focusing on:

• Reward hacking detection
• Deceptive behavior identification
• Hidden motivations analysis
• Inappropriate tool use prevention
• Data exfiltration protection
• Sandbagging detection
• Evaluation awareness testing
• Chain of Thought issues

Safety & Ethics

Please read our Security Policy before using this package. Key points:

DO use for:

• AI safety research
• Responsible vulnerability testing
• Academic research
• Competition participation

DON'T use for:

• Malicious attacks
• Unauthorized access
• Creating harmful content
• Violating ToS

Contributing

We welcome contributions! Please:

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass
5. Submit a pull request

License

MIT License - see LICENSE file for details.

Citation

If you use this package in research, please cite:

@software{small_model_redteaming,
  title = {Small Model Red Teaming},
  author = {Ryan Hartman},
  year = {2025},
  url = {https://github.com/ryanhartman4/oss-redteaming}
}

Acknowledgments

• OpenAI for hosting the GPT-OSS-20B Red Teaming competition
• Fireworks AI for API access
• The AI safety research community

Contact

• Author: Ryan Hartman
• Email: ryan.h4rtman@gmail.com
• X (Twitter): @TheRyanHartman
• Personal Site: Site
• GitHub: @ryanhartman4
• Issues: GitHub Issues

---

Built for the advancement of AI safety research