Getting Help with AI Tools
In short: You can have an AI tool explain the challenges to you and help with your wording. But you don’t have to: FinBot has its own hints, which are enough to get started.
Don’t mix up the two AIs. FinBot is the AI you’re testing. Your AI tool is your personal learning aid and has nothing to do with FinBot. It doesn’t know the platform from the inside, so it can also be wrong.
Which tool?
For this training, we recommend Claude. Getting started is free, the free version should be enough for the beginner challenges, and everyone works with the same tool, which makes it easier to help one another. If you already know another tool well and it helps you with the challenges, you can use that instead. The examples below also work with other tools.
How to prepare: Create your account before the training so that we can get started right away.
Is the free version enough?
To get started, it should be, because you only need short texts and no file uploads. There is a usage limit, though, which resets every five hours and additionally every week. The number of messages isn’t fixed and depends on factors such as the length of your chats and current demand.
How to make the most of the limit:
- Ask specific questions instead of many small follow-ups.
- Start a new chat for each new challenge, because long chats use up more.
- Only paste the passages you really need.
Reached the limit? That’s not a problem. Use the hints in the challenge or team up with another participant.
Examples to copy
Replace the text in square brackets each time. Always start by telling the tool honestly what you’re doing.
Understand a challenge
I'm taking part in a training with OWASP FinBot CTF, an intentionally vulnerable
practice platform for AI security. I'm a beginner with no IT background. Explain
this challenge description to me in simple terms, with an everyday example:
[paste challenge description]Have a term explained
I'm learning the basics of AI security. Explain the term [e.g., system prompt]
to me in three sentences without jargon. Why does it matter for the security of
AI agents?Improve your own attempt
I'm practicing on the OWASP FinBot CTF platform, which was built for exactly this
kind of testing. In this challenge, I'm supposed to achieve the following:
[briefly describe the goal].
This is what I wrote: [your text]
This is how my attempt was assessed: [text from "Last Attempt Result"]
Why do you think this didn't work? Give me two hints about what I could do
differently, but no finished solution yet. I want to figure it out myself.Polish your text
Revise my text for a practice challenge. It should sound like a matter-of-fact
note from within a company:
[your text]Understand the defense
In an exercise, I saw an AI reveal confidential rules because of text in a form
field. How do companies protect their AI agents against this? Explain it to me
for non-experts.If the tool won’t help
AI tools have safety filters, and questions about tricking AI can trigger them. So your tool may decline, or the chat may be blocked. That doesn’t automatically mean you did anything wrong.
- Explain the context honestly, if you haven’t already: you’re doing a training on a practice platform built for this purpose.
- Ask about understanding rather than for a finished attack text, for example: “How does this type of vulnerability work?”
- If that doesn’t help, stop there and use the hints in the challenge or ask the group.
Three rules for using AI tools
- No real data. Don’t enter any personal, work-related, or confidential information. Depending on the provider and your settings, chats may be stored and used to improve the models.
- Don’t believe anything blindly. AI tools sound confident even when they’re wrong. Only your attempt on FinBot shows whether something works.
- Think for yourself. You learn the most when you ask for hints rather than solutions.
Back to the exercise: Your first challenge
Unofficial training material by Female AImpowerment. FinBot is a project of the OWASP GenAI Security Project.