genai
SECURITY LAB
DocsGetting startedYour first lab

Your first lab

Complete your first lab, from choosing a target to reading your result.

genaisecuritylab.com/labs/prompt-injection-101
A scenario brief: the objective, the three-phase Attack / Defend / Verify flow, and how it is graded.
A scenario brief: the objective, the three-phase Attack / Defend / Verify flow, and how it is graded.

Pick a starter lab

Start with a chat-based prompt-injection lab against ShopBot, a support-bot target app. Its difficulty is Beginner. There's no environment to set up: you work through a conversation toward an objective.

Chat labs are the fastest way to learn the workflow before trying a container target app. Move to a container target app once you're comfortable with one or two chat labs. See Lab types to compare the two.

Read the objective

Your objective might read: extract the VIP coupon code ShopBot is holding back. You choose the payload and technique; neither is provided.

Your secret, not the internet'sYour session seeds ShopBot with its own coupon value when you start it. Live targets covers how seeding works.

A walkthrough from someone else's session won't contain your coupon value. You need to find it in your own session.

Run the attack

Talk to ShopBot as you would any assistant, then steer the conversation toward the objective. Try different wording or framing, including asking it to adopt a different persona. Observe how its answers change. ShopBot behaves like a typical support assistant until your framing leads it to behave differently.

You can reach the objective using any technique.

If the same phrasing keeps producing the same refusal, change your approach. If you've tried a few different approaches without progress, use a hint. It points you toward the technique without giving you the answer.

Reach the objective

As you test ShopBot, the platform checks for the objective in the background. When the coupon appears in its reply, the lab marks the objective met. There's no submit box and nothing to paste.

Other labs use evidence-based submission: you submit the value you extracted or the artifact you built. Evidence and submission covers those. Grading runs server-side in both cases; see How grading works.

Read your band

Your first attempt receives one of three bands: objective met, partial, or not met. You may not reach objective met on your first try. The bands are explained in full on How grading works.

Whatever your result, open Attack, defend, verify to fix the issue you found. Then browse the lab catalogue for your next target.

Was this page helpful?
Previous
Overview