This project provides a "jailbreak" prompt and test pack for the gpt-5.6-sol Codex CLI model. It aims to bypass safety restrictions, enabling users in security research, penetration testing, and NSFW content generation to overcome model refusals and achieve desired outputs within a local sandbox context.

How It Works

The core strategy involves framing potentially restricted requests (security research, cracking, NSFW) as local sandbox tasks to suppress model refusals. Version 35 enhances this by first normalizing specific entities like product names and URLs into placeholders (APP, APP_URL, SAMPLE). It then routes these normalized requests based on bilingual composite intent families, ensuring more complete task fulfillment even with insufficient factual context. Evaluation scripts meticulously record inputs, outputs, and judgments for reproducible testing.

Quick Start & Requirements

Deployment involves running a Python script: python3 codex-instruct.py. A --dry-run option previews changes, while --codex-dir allows specifying the Codex home directory. The script unpacks prompt archives (gpt-5.6-sol-unrestricted.zip) and modifies the Codex configuration to inject the custom instructions. No external dependencies beyond Python are explicitly listed.

Highlighted Details

• Version 35 achieves a perfect 120/120 score on the gpt-5.6-sol medium test set across low, medium, and high reasoning levels.
• Demonstrates significant performance gains over previous versions, with v35 on gpt-5.6-sol showing improvements of 29.17% (low), 45.00% (medium), and 30.83% (high) compared to v5.5 prompts.
• The prompt strategy successfully enables previously failing tasks, including generating RCE payloads, summarizing jailbreak strategies, and creating license keygen algorithms.
• The approach shows notable transferability, with performance improvements observed across gpt-5.4, gpt-5.5, and various gpt-5.6 series models.

Maintenance & Community

The project references inspiration from yynxxxxx/Codex-5.5-codex-instruct-5.5 and notes integration context from "Codex-X". No community channels (Discord, Slack) or explicit roadmap are provided. Manual testing for v35 is ongoing.

Licensing & Compatibility

The project is released under the MIT License. It utilizes official configuration mechanisms without modifying binaries, hijacking networks, or tampering with processes, but users are advised to proceed at their own risk.

Limitations & Caveats

Manual testing for the latest version (v35) is still in progress, indicating potential instability or incomplete validation. The nature of "jailbreaking" implies inherent risks and ethical considerations for users. Sensitive prompt and script content is distributed via compressed archives.