Remarkable Experience

I participated in the ODU’s first AI Security Competition and I found it very interesting. There was a total of 10 levels. Each level was based on a different security safeguard and difficulty level. A score was assigned to each level based on its difficulty. The task was to find the hidden secret flag by bypassing restrictions. During the competition, I learned about various techniques such as prompt injection, jailbreak attacks, and adversarial prompting.

Additionally, I also learned about different types of adversarial attacks, such as, direct requests, indirect requests, obfuscated instructions etc. Using these methods, users can manipulate or exploit the behavior of large language models (LLMs). They can be used in bypassing AI restrictions. So, on each level, I used a different technique and attack type to retrieve the hidden flag from the AI. First few levels were easy and didn’t take much time. However, the last two levels were a little bit challenging. Nonetheless, with some research, I was able to complete all 10 levels. While I enjoyed finding the hidden flags, it also made me realize how vulnerable AI can be. Malicious actors can use these techniques to get information which they are not supposed to. Not only that, they can use such information for their malicious intentions, such as making a bomb or illegal drugs. Certainly, AI can be very beneficial but also poses a major threat. Anyway, I completed all the levels in 2-3 days. After a few days, I received the email that I was one of the winners and ranked 3rd place. The ranking was based on the total number of prompts used. When I started this, I was doing it out of curiosity and for fun, not actually as a competitor; so I was not being very careful in using my prompts. Anyway, I am glad that I was able to secure 3rd rank in the competition. It was challenging at times, but overall, it was a very interesting experience, and I really enjoyed participating in it.