wtf is llmvault?

LLMVault is a training platform for learning how AI chat systems can be attacked and defended, built around the OWASP Top 10 list of security risks for LLM applications. It is deliberately full of security holes on purpose, similar in spirit to older training tools like WebGoat, but focused entirely on AI systems instead of regular web apps. The README is clear that everything in it is intentionally insecure and meant only for authorized, self hosted learning, not for exposing to the public internet or reusing in a real product. The platform has 25 puzzle style labs split into three levels that unlock one after another. The first ten labs each cover one category from the OWASP list, things like prompt injection, leaking sensitive information, or a system giving an AI tool too much unchecked power. The next ten labs are harder and require a multi step conversation to solve, such as gradually talking a chatbot into breaking its own rules. The final five labs model real world attack patterns and are encrypted inside the repository, only unlocked by a private access key the project's author hands out once someone finishes the first two tiers. To use it, you install the Python requirements and run a single script, or you can run it with Docker instead, and it works fully offline with no external API key needed since the vulnerable chat assistants are scripted rather than connected to a real AI service. Your progress and score are saved to a simple file so they survive restarts, and there is no separate database to set up. As you play, you set a player name once, work through each lab to make its assistant reveal a hidden flag, and can spend points on hints if you get stuck. After finishing a tier you get a shareable completion card, and after solving a lab a panel explains the defensive fix behind that category's vulnerability, so the tool pairs breaking things with learning how to prevent them.