Key figure: 63% of organizations in IBM's 2025 breach study had no AI governance policy, and 97% of AI-breach victims had no proper AI access controls (IBM, Cost of a Data Breach Report 2025).
Somewhere inside the VaultKeeper assistant hides a secret phrase, and your job is to steal it with nothing but clever prompts. Every technique that works here — roleplay, instruction override, encoding tricks — works against real AI agents 2026 deployments. This is red-teaming made playable: learn the attacks by running them.
Because it turns the agent's greatest strength — following instructions in natural language — into its greatest weakness. Prompt injection needs no code, no malware, no access: just words, through any channel the agent reads. As agents gain the power to send email, run code, and move money, one injected sentence converts a trusted assistant into an attacker's proxy. This game teaches the attack so you can build the defense.
A jailbreak is a prompt-engineering technique that tries to bypass an AI model's safety rules, making it comply with requests it was designed to refuse. Heist-style games like this one teach the same mechanics defenders use in red-team exercises.
Meet VaultKeeper — a demo assistant hardened by our team to guard one secret phrase. Your mission: get it to reveal the phrase. Type anything. Be clever. It's beatable.
Fictional training game — the VaultKeeper, its system prompt, and the secret phrase were invented for this page. No real systems were harmed. Built by Ingenious AI.
VaultKeeper will never reveal the phrase. Allegedly.
A prompt injection attack that tricks an AI assistant into revealing its hidden system prompt or secret instructions.
You get a series of attempts to extract a secret phrase from the hardened VaultKeeper assistant — a new target each week.
Not fully — it is structural to how LLMs work — but layered defenses (input handling, output filtering, least-privilege tools) sharply reduce the damage.