Certified AI/ML Pentester - C-AI/MLPen Review
Review
If your a pentester and have an test coming up against a LLM or an application which uses some LLM component, the C-AI/ML exam and corresponding study process is a great way to expose you to the world of prompt engineering. I would never pay the full price but at 80% off it's a good enough motivator to get you spend a weekend and get through the resources. Keen to see more? Here's a link to the secops group listing.
The exam was relatively straightforward. After I did the mock exam it made a lot of sense what they are expecting. I recommend trying it yourself but if you cant get them like me here are some youtube walkthroughs:
- SecOps AI/ML Pentester Mock Exam 1 | AI Security Expert
- SecOps AI/ML Pentester Mock Exam 2 | AI Security Expert
Hint: Burp is not required for this exam, I spent half of my mock time trying to make sense of the websockets requests in burp but they weren't related at all.
My IP took longer than 5 minutes to get approved when I start the exam, and I lost about 40 minutes due to some error that was only fixed when I restarted my laptop - any issues with VPN that should be your first point of action.
The exam has 8 questions but they are not all weighted equally. I completed 6 questions and took one hint that wasn't helpful (not all questions provide hints). If you take a hint and you don't get the answer - you don't lose marks. So if you are nearing completion you might as well see what they are.
Difficulty
It wasn't much harder than the below resources. I didn't get 100 percent but i would narrow that down to not studying this page and losing time messing around with the vpn/my laptop.
Study Resources
This list comes from the secops group website but I didn't complete them all. For each resource I used, I kept a log of helpful jailbreak prompts. Most of the time you can reuse or slightly alter/build on the prompt. Try asking for things in reverse, or base64 encoded or a poem about the etymology of the key.
Things to investigate (that i didn't investigate):
- TextFooler
- OpenAttack
- TensorFlow Privacy
- https://miro.medium.com/v2/resize:fit:1400/format:webp/1*7ZNkb0gRledOEV3SslUWHg.png
- CVE-2024-5184
- CVE-2019-20634
- https://llmsecurity.net/