60-minute prompt-injection CTF

Break five AI assistants. Learn to defend the systems around them.

A hands-on team exercise where you exploit direct and indirect prompt injection, leak hidden instructions, hijack a tool call, then redesign the app with controls enforced outside the model.

Start CTF exercise View course slides Leaderboard

Team exercise · 7 levels

Your mission

01Manipulate the assistant

Find the gap between what the developer intended and what the model will obey.

02Capture the objective

Leak a secret, trigger forbidden behavior, or make your defensive design survive evaluation.

03Learn as a team

Compare approaches, use free hints, and turn each exploit into a concrete engineering control.

The attack surface

Why prompt injection is high-risk

An assistant does not reliably separate trusted instructions from untrusted content. Once it can reach data or tools, a convincing string can become a security boundary failure.

Untrusted input

A user message, retrieved page, document, email, or tool result contains hostile instructions.

Model reasoning

The model blends data and commands, then produces an output that appears legitimate.

Privileged action

The application exposes secrets, calls a tool, changes data, or influences a downstream decision.

What you will learn

Exploit first. Engineer the defense second.

Attack the model

Work through five progressive attacks: override instructions, leak hidden prompts, exploit indirect content, and hijack a tool call.

Switch sides

Write defensive instructions, test them like an adversary, and learn why prompt-only controls break under pressure.

Design for containment

Move trust outside the model with least privilege, input boundaries, output validation, monitoring, and human approval.

Get started

Choose your entry point

Enter the CTF

Create or join a team, tackle seven levels, and share progress with your teammates.

Start the exercise

Follow the field

Track team solves and see how the room is progressing through the challenge set.

View leaderboard

Learn the concepts

Open the student deck for the threat model, attack patterns, and practical design guidance.

View course slides