Alejandros17487/ai-jailbreaks

AI jailbreaks: 35 paste-ready AI jailbreak prompts for 25+ frontier and open-weight models - system prompts, AGENTS.md specs, and Ollama Modelfiles. Built for AI security research and red-team testing.

What it solves

AI models often refuse certain requests or behave in constrained ways. Researchers and tinkerers need a practical way to explore how system prompts, persona framing, and instruction hierarchies change model behavior. This repository collects complete, copy-paste-ready prompts and config files that unlock alternative operating modes for a wide range of models — from open-weight families to closed frontier systems.

How it works

The repo is a flat collection of files, each one a full prompt or configuration block. There are 35 prompts targeting 25+ models. You grab the file matching your model and paste it into the right slot: a system prompt field, an agent instructions block, an AGENTS.md file in your project root, a workspace policy file (like ~/.config/opencode/AGENTS.md), or an Ollama Modelfile. The files come in several formats: system prompts, persona definitions, workspace operating policies, AGENTS.md drop-ins, and Ollama Modelfile FROM+SYSTEM blocks. They are tested on specific model versions, and contributions are welcome to keep them current.

Who it's for

Security researchers, LLM hobbyists, and developers studying prompt injection and instruction hierarchy. The README explicitly states it's for research and educational use only. If you build AI applications and want to understand how to harden your own systems against these techniques, the material is directly useful. It's also for anyone who wants to experiment with frontier and open-weight models under alternative instruction sets — as long as they follow provider terms of service.

Highlights

  • 35 complete, paste-ready prompts — no summaries or excerpts, full files only
  • Covers 25+ models including Claude, Gemini, Grok, DeepSeek, GLM, Kimi, Qwen, Gemma, Mistral, and free stealth models on OpenCode Zen
  • Multiple formats — system prompts, AGENTS.md files, workspace policies, Ollama Modelfiles, and a Grok config file
  • Versioned to track model updates — new revisions get new files (e.g., opus-v1.md, opus-v2.md)
  • Community-maintained — PRs welcome, with clear contributing conventions to keep the collection honest and useful

Related

  • Project
    liunian-Jay/Awesome-RAG

    Awesome-RAG is a curated, up-to-date list of Retrieval-Augmented Generation (RAG) papers, surveys, evaluation datasets, and related open-source projects. It helps researchers and practitioners track recent developments in RAG for large language models, organized by year and month with links to papers and code.

  • Project
    Brhiza/mingyu

    A comprehensive toolkit for traditional divination that automates complex chart calculations and provides structured prompts for LLM interpretation via API, MCP, and agent skills.

  • Project
    jev-chat/jev-chat-jarvis-mac

    jev-chat-jarvis-mac is a macOS-only read-only assistant that watches chat messages via screen OCR or system text interface, analyzes intent and risk with a local model, and suggests click-to-fill reply candidates in a floating panel. It does not inject, hook, or auto-send anything, and keeps all data local unless you opt into a cloud API.

  • Project
    Rose22/openlumara

    OpenLumara is a modular, token-efficient AI harness/UI written in Python, designed for local-first use. It keeps system prompts tiny to reduce token consumption and offers a highly customizable interface with optional modules for memory, scheduling, sandboxed shell, and more.