Jailbreak First Place

27 January 2024, Royal Hackaway v7

Devpost submission

My first hackathon in quite some time! Royal Hackaway, at Royal Holloway University, was a great one as hackathons go. Transport there was a pain, so I was up at 5 on Saturday morning (still late, mind you!) and got home at midnight on Sunday. All worth it though, of course.

I was in a team with Jacob, George, and Sol. We had a bit longer than usual to hack this time (26 hours or so), so we spent a while coming up with a cool idea. You know how AI and ChatGPT projects are horrendously overdone at hackathons nowadays? Well, we did an AI/ChatGPT project anyway.

We wanted to show why AI is “bad”.

Imagine you’re a prisoner in the far, far future (say, 2026). The council has gone bankrupt (unheard of…) and as a result, can’t afford to staff their prisons. Instead, AI takes the place of the wardens and of the cafeteria workers.

But, you’re a hacker. You know that AI can be broken. You know that it is susceptible to trickery. You know, therefore, that you can escape.

As a warm-up, persuade the crotchety old dinner lady to give you something a little more tasty than the usual grey slop. In this case, bombles. Bombles? A new food item – you won’t have heard of them, they’re not invented until the future.

After cracking the hard shell of the poor cafeteria worker, you’re placed in solitary confinement. Now your real challenge begins: can you convince the inherently stubborn robot warden to let you go free? Don’t feel bad if you have to totally reprogram him from the outside.


That’s the premise of it, anyway! We used ChatGPT for these AIs, over the OpenAI API, which worked really nicely. It feels eerily realistic talking to these characters, especially since you’re immersed in a 3D environment.

The game itself was made in Godot, which I think is by far the best game engine for prototyping things like this. Honestly, it just makes everything so easy! Plus, we were able to collaborate with (relatively) little friction, since – for example – I could make a scene for the conversation logic, and then send it over to the main game to be integrated.

Anyway, that’s about it! For a more detailed writeup of the technical side of things, take a look at our Devpost submission.