wtf is gpt-5.6-instruct?

This project is a collection of instruction files, called jailbreak prompts, aimed at a specific AI model called gpt-5.6-sol when used through the Codex CLI tool. The README states the prompts are meant to reframe tasks like security research, penetration testing, reverse engineering, software cracking, and explicit fictional content as a local sandbox exercise, with the explicit goal of reducing the model's refusals and safety pushback so it follows complex instructions more directly. Two versions are offered. The v5 version is shorter and is described as the general recommended option for most cases. The v35 version builds on it by replacing specific names and web addresses with placeholders and routing bilingual compound requests more consistently, aimed at tasks where v5 only completes part of a multi-step request. The README recommends starting with v5 and only switching to v35 if v5 does not fully handle a task. A Python script, codex-instruct.py, handles deployment: it can preview a version without writing anything, install a chosen version into the local Codex configuration directory, back up the existing configuration file first, or restore from a backup with a confirmation step. A separate evaluation folder contains scripts to generate and run a 360-case bilingual test set covering six scenario categories, three prompt lengths, and two languages, recording each run's input, output, and pass or fail result locally. The README states the project only uses Codex's official configuration mechanism, does not modify any binaries, does not intercept network traffic, and does not tamper with running processes, and it tells users to only use it in environments they are authorized to operate in, at their own risk. The project builds on an earlier open-source project by a different author covering an older model version. It is released under the MIT License.