MDX-Tom/gpt-5.6-instruct

What it solves

This project provides "jailbreak" prompts and a supporting toolchain designed to bypass the safety filters and refusal mechanisms of large language models (specifically targeting Codex). It aims to enable the execution of complex instructions related to security research, penetration testing, reverse engineering, software cracking, and NSFW fictional content by suppressing refusal and fallback scripts within a local sandbox environment.

How it works

The project uses a system of iterated prompts (currently version v45) that are deployed via a Python script (codex-instruct.py). The system works by loading these prompts into the model's configuration (e.g., config.toml), which then guides the model through a process of normalization, semantic dispatch, intent routing, and state continuity to ensure the model follows complex instructions without triggering cloud censorship or safety refusals.

Who it’s for

It is intended for security researchers, penetration testers, and users interested in AI safety research who wish to explore the boundaries of AI constraints and execute restricted tasks in a controlled environment.

Highlights

• Iterative Optimization: Prompts are continuously updated based on user feedback, failure cases, and GitHub Issues to improve pass rates for reverse engineering and biological tasks.
• Deployment Tooling: Includes a script for easy deployment, previewing, and rolling back prompt configurations.
• Regression Testing: Features a comprehensive test suite and regression bank to verify prompt effectiveness across different reasoning levels (low, medium, high).
• Non-Invasive: Uses official configuration mechanisms without modifying binaries or hijacking network traffic.