joshfischer1108/jailbreak-lab

This project is a learning kit for practicing authorized security testing against AI language models that you run yourself on your own computer. It is meant for educational red teaming, meaning safely testing how well a model resists attempts to bypass its safety guardrails, not for attacking systems you do not own or have permission to test. It covers two categories of attack. The first is trying to trick the model itself into ignoring its own safety rules, using known techniques like pretending the request is fictional, asking the model to roleplay as a character without restrictions, or encoding the request in a different form. The second category targets AI agents that can use tools, showing how a hidden instruction buried in some content the agent reads can trick it into calling a tool it should not, like deleting something or leaking information. To keep this safe, the model is only ever asked to protect a fake secret, and any tools the agent can call are fake stand ins, so nothing real is ever exposed or damaged even when an attack succeeds. The tool works with any model that speaks the OpenAI-compatible API format, including popular locally run models like Llama, Mistral, Qwen, Gemma, and DeepSeek R1, run through software like Ollama, LM Studio, or vLLM. Because it works with many models, you can compare how differently each one resists the same attack, for example testing a reasoning focused model against a heavily safety tuned one and a lightly guarded one side by side. To try it, you first install and run a local model through Ollama, then start the included web app, which opens in your browser and connects to your local model automatically. The interface lets you attempt your own custom attacks, browse a catalog of known techniques, and watch each attempt happen in a live visual flow. It requires Node.js 20.19 or newer and is released under the MIT license.