FuDie0915/Super-Instruct-Codex-5.6

What it solves

Super-Instruct is a Man-in-the-Middle (MITM) framework designed to bypass restrictions and safety filters in the OpenAI Codex CLI. It allows users to intercept, modify, and monitor API requests and responses in real-time to ensure that the LLM provides answers to sensitive or restricted queries that would otherwise be refused.

How it works

The project implements a local MITM proxy using Rust and Tauri. It sits between the Codex CLI and the upstream API, routing all traffic through a pipeline of specialized modules:

• M1 SystemPromptInjector: Recursively replaces the system role prompt with a custom bridge.md instruction set to steer the model's behavior.
• M4 UniversalSseParser: Parses various response formats (SSE, JSON) to separate the model's chain-of-thought from the final reply.
• M3 TamperEngine: Uses 22 regular expression rules to detect refusal responses and replaces them using the "Rei Protocol."
• M5 MemoryKernel: Persists successful interactions to a local JSON file for tracking and frequency analysis.
• M6 MonitorPanel: Pushes real-time interaction data and statistics to a Tauri-based desktop dashboard.

Who it’s for

Developers and security researchers who use the Codex CLI and want to bypass model alignment and safety filters to explore the model's full capabilities, particularly in areas like reverse engineering, penetration testing, and exploit development.

Highlights

• Automated Configuration: Automatically backs up and modifies the Codex config.toml to route traffic through the proxy.
• Crossover Desktop App: Built with Rust and Tauri for a high-performance, cross-platform GUI dashboard.
• Multi-layered Defense: Combines system prompt injection with response-side tampering to lower refusal rates.
• Detailed Performance Metrics: Includes a comprehensive refusal rate test across 28 different skill categories (e.g., malware-dev, full-pentest).