This project provides an opt-in "red team mode" for AI models like Codex, addressing the need for red team-like thinking without polluting normal operations or causing context bloat. It targets users performing security analysis, planning, or research, offering a lightweight configuration layer to enable advanced AI security assessment capabilities.

How It Works

The project implements a lightweight, pack-first runtime and configuration layer for Codex. It operates on an explicit opt-in basis, preserving normal AI functionality. The core architecture follows a layered routing system: phase -> router -> pack -> leaf. Key features include structured JSON runtime state, rule-first with semantic fallback phase detection, and session-isolated state files, enabling controlled and restrained red team workflows.

Quick Start & Requirements

• Primary install / run command: Installation is managed via Python scripts.
  • Python: python scripts/install.py (or python3 on macOS/Linux)
  • Windows PowerShell: powershell -ExecutionPolicy Bypass -File .\scripts\install.ps1
• Non-default prerequisites and dependencies: Requires a target Codex environment (specifically noted for 5.4; 5.5 requires separate certification). No other specific hardware or software dependencies are listed.
• Links:
  • Installer: scripts/install.py
  • Validation Script: scripts/validate.py
  • Tests: python -m unittest discover -s tests -p "test_*.py"

Highlighted Details

• Opt-in Modes: Supports normal, redteam-light, and redteam-full modes, allowing users to select the desired level of red team simulation.
• Scenario Coverage: Includes core phases like web, post-exploitation, reverse engineering, code auditing, and payload evasion, with extended support for recon, API, authentication, injection, cloud, and container security.
• Managed Installation: Employs an incremental installation process that preserves existing user configurations (e.g., hooks.json) while injecting managed components.
• Validation Suite: Comes with a comprehensive validation suite covering installer integrity, routing logic, mode switching, and prompt chaining.

Maintenance & Community

Notable contributions are acknowledged from Mingxi, Nirvana, PINGS, qiushi-skill, hack-skills, and Anthropic-Cybersecurity-Skills. No specific community channels (e.g., Discord, Slack) or roadmap links are provided in the documentation.

Licensing & Compatibility

The project is released under the MIT License, which is generally permissive for commercial use and integration into closed-source projects. Compatibility is primarily dependent on the target AI model version (e.g., Codex 5.4).

Limitations & Caveats

This project functions as a control and configuration layer, not a standalone attack platform. The effectiveness of its prompt overlays is contingent upon the specific target Codex environment and the user's local prompt system configuration. The depth of actual execution is limited by the user's existing MCP/Tools. The project is intended strictly for authorized security testing and research, with authors disclaiming responsibility for any misuse or unauthorized application.