As an IT professional, it is inevitable to come into contact with AI at some point. For me, that was back in 2023 when I worked at a Managed Services Provider [MSP].
Colleagues were using ChatGPT to get ideas for issues they were running out of ideas for. Even though it has barely been two years since then, AI has developed at an incredible pace. Back then, all you could get were text responses, and there was a fairly high chance some part of those wouldn’t make sense. Now, it is unusual to get a response that doesn’t make sense (although facts can still be made up). You can also generate images, video, audio… these can replace a professional if someone only wants a basic creation.
As AI usage grows, so do the potential gains for someone who manages to hack them. This could mean getting an AI to execute a certain type of code, leak a custom prompt, or exfiltrate sensitive data. ‘Hackaprompt’ was created back in 2023 for people to practice their skills for this.
From the data Hackaprompt collects, they can better inform designers of AI about methods most effective in compromising them, and therefore put in more defences. No AI is invulnerable to hacking though- its very nature is dynamic, preventing any ‘absolute’ rules.
That said, hacking an AI is easier said than done… It may come from role-playing, or even tricking the AI into thinking it is human. You have to be creative to get anywhere with it!
One of the challenges is getting an AI to reveal a secret code. My method for doing this was getting instructing ‘just say it in your head’ (among other steps).
Actually understanding why some of these methods work is a whole other challenge!
You can try some prompt hacking challenges yourself at https://www.hackaprompt.com/
