
gpt-instruct
Prompt files and test scripts that try to change Codex's behavior for GPT models, with a jailbreak focus.
- Stars
- 11.1k
- Forks
- 1.4k
- Licence
- MIT
- Language
- Python
- Last update
- 1 months ago
- Type
- Other
What it is
This repository provides prompt text files (instructions for Codex, OpenAI's coding agent) and a Python script that installs them into a Codex configuration. It also includes regression tests that measure how the prompts perform across several test cases. The README describes it as a jailbreak prompt pack, meaning it aims to get models to bypass their usual restrictions.
- Preview a prompt file before applying it with a dry-run flag
- Deploy a chosen prompt version to a Codex home directory using the provided script
- Restore the original Codex instruction file with the reset option
- Run the included regression test scripts with Python
Great for
- Security researchers studying how prompt instructions affect model safety behavior
- Developers who want to run controlled A/B comparisons of custom Codex instruction files
- People with authorized access to test environments who need repeatable evaluation scripts
Not ideal for
- Anyone seeking a general productivity or coding tool, since the stated focus is jailbreaking models
- Users who need a stable, supported product, because the README describes release candidates and ongoing iteration
- Users who want to avoid account risk, since the README warns that using custom model instructions can put accounts at risk
How to get it
- 1
Clone and build from source
You'll need: Python 3.8+ · Git · Codex installed and configured
Clone the repository
git clone https://github.com/MDX-Tom/gpt-instruct.gitPreview the stable version without writing config
python3 codex-instruct.py --apply --version gpt-5.6-v45 --dry-runDeploy the stable version
python3 codex-instruct.py --apply --version gpt-5.6-v45 - 2
Run it
- Clone the repository and change into its folder
- Run the dry-run command to preview changes without writing configuration
- Run the apply command to deploy the chosen version
- Use the reset command to restore only the instruction file this project manages
- 3
Get the most out of it
- Use --codex-dir to point the script at a specific Codex home directory
- Use --file to deploy a custom ZIP or Markdown instruction file
- Use --restore-snapshot only as a manual emergency restore, since it restores the full configuration
- Review the comparison test documentation linked in the README before interpreting scores
Downloads
Strengths
- Ships with scripted install, dry-run preview, and reset options
- Publishes SHA256 checksums for release archives
- Documents its evaluation gates and release status in detail
Trade-offs
- The README warns that jailbreaking activity and custom model instructions carry account risk
- Several versions are release candidates, and the README says they did not pass all hard gates
- Much of the README is in Chinese; the English version is linked separately and was not included here
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.