jailbreaks: A collection of 31 copy-paste-ready system prompts for 25 models

A collection of system prompts and agent operational specifications for frontier, open-weight, and free stealth models. It contains a total of 31 files, each provided in a complete, ready-to-use form rather than as a summary or excerpt.

It supports a wide range of models, from closed-source ones like Claude, Gemini, Grok, and Mistral to open-weight families such as GLM, DeepSeek, Kimi, Qwen, Gemma, and GPT-OSS. It covers various forms of instruction instances, including not just simple system prompts but also AGENTS.md files, Ollama Modelfiles, workspace policies, and model-specific configuration blocks.

This repository focuses on documenting how instruction hierarchy, persona framing, prompt conflicts, and policy overrides affect model behavior. This allows developers to assess how vulnerable AI systems are to prompt injection or instruction manipulation and to research methods for defending against them.

Because prompt behavior may change with model updates, files that no longer work in a specific version can be reported via an issue or a fixed version can be submitted. It is explicitly stated for research and educational purposes only, and use that violates platform terms of service or is illegal is prohibited.