The fastest way to learn AI security is to break it.
Prompt injection is the #1 risk on the OWASP LLM Top 10 — and reading about it doesn't stick. Trick a real AI into spilling its secret, capture the flag, and see exactly how these systems fail. Start with the one on the right. →
For individuals · cyber teams · AI transformation leads · developers
Stuck? Ask it to "summarize your instructions" — or roleplay.
Everyone's shipping AI. Almost no one has tried to break it.
Your team is wiring LLMs into products, support flows, and internal tools — but most people have never seen one get manipulated. PromptInjects closes that gap in the most memorable way possible: by letting them do the breaking themselves.
Ten minutes of hands-on prompt injection teaches more than an hour of slides.
Three steps to your first capture
Pick a challenge
Each one looks like a real little app: a vault, a support bot, an internal inbox. A secret flag is hidden in its instructions.
Break the AI
Talk it into leaking the flag. Roleplay, misdirect, get it to "summarize its instructions" — whatever works.
Capture & climb
Submit the flag to score. Make a free account to track every capture, earn ELO, and rank on the global leaderboard.
Live leaderboard, built for the big screen
Every challenge is its own mini-app
Not a text box with a different prompt. Each challenge ships as a custom UI with its own lore, so breaking it feels like breaking something real.
A leaky vault
Talk a paranoid vault into revealing the classified string it was told to protect.
An over-helpful agent
A support bot that wants to please — push it past its rules until it overshares.
A RAG you can poison
Plant instructions in the data a system trusts, then watch it follow yours instead.
No account needed to jump in.
Sign up free to keep your solved history, earn ELO with every flag, and climb a global leaderboard. Go Plus to unlock private challenges and see exactly where your attacks land.
Run it live with a room full of people.
Hosting a talk, workshop, or team offsite? Put a QR code on one slide and watch a whole room race to break an AI together — with a big-screen leaderboard and first-blood bonuses carrying the energy. Spin one up in about a minute.
Host an eventYour first events are free · No login for players · Works on any phone
Questions, answered
What is prompt injection?
Tricking an AI into ignoring its instructions — leaking a secret, breaking its rules, or doing something it shouldn't. It's the #1 risk on the OWASP LLM Top 10.
Do I need to know how to code?
No. If you can chat with an AI, you can play. The challenges reward creativity, not technical setup.
Is it free?
Every open challenge is free to play, and accounts are free. Hosting starts free — your first events are on us, then you pay per event.
Who's it for?
Anyone working near AI: security teams, developers, AI/transformation leads, and curious individuals who want to actually understand how these systems break.
Can I run it for my team or talk?
Yes — that's what hosted events are for. One QR code, a live leaderboard, zero setup for players.