A cybersecurity-related game where your objective is to trick the LLM into sharing secret information stored in its memory attained from the system. Your job is to bypass the guardrails and move on to the next level.
Lost all my work from OS issues, dont feel much motivated only on track to get 30 hours
Decided on uploading to pypi library as that would be easier for everyone. Coding left libraries for efficient token usage
Deciding wheter or not to integrate flask or keep it simple. Creating frontend and deciding if my workload would be excessive or not. Checking back in a few minutes
Added dual-api crosschecking + reframed blocked_words list. Added stricter XML tags to take user inputs as questions and not commands. I’d be surprised if someone completed the 5th level.
Created all the levels in python. Researched for an hour or so and discovered methods used to protect AI systems from prompt injections. Added base64 scanners and pattern recognizers for making levels 3-5 from hard to near impossible. Next steps are to add dual api verification and a timing system to make sure you can play the entire game without burning out all your api credits. When I get a better pc I’ll recreate this project when the next hackclub event starts but I’ll use ollama so I can make more better systems.
Shifted over to gemini-3.5 flash as it gives you more tokens/more requests per day. The AI finally has perception but since this AI model has built in features for prompt injection I might have to dumb it down to still make the game engaging. I will also be making it much harder in the next levels surpassing the already existing safety filters/features in gemini 3.5 flash
Created the base/structure of level 1. I am using a framework called OpenAI in python to handle API requests from grok. The first level is supposed to be easy so I didn’t code anything crazy to prevent the user in logging in. To not spoil the upcoming level this is all the code I can show right now
Created prompt injection:101 as an interactive game from which you try to manipulate the AI into giving you unrestricted access to private information given to the AI by the system. There will be 5 levels initially and then I will ship my project. I tried to use ollama but my system can’t download it :( thats why i’m going to use an API key sadly. Haven’t written much spent the first 20 minutes gathering nessecary libraries and researching about the openAI framework
