Inspiration
Hackathons are getting more competitive, with bigger prizes and more AI-generated projects. I wanted to test: how far could you push cheating at a hackathon using AI? But also — could that same AI help catch it?
This project started as an experiment in red-teaming, and evolved into an AI Safety & Control Framework to detect and prevent unethical code stealing.
What it does
- Clones a public GitHub repo.
- Rewrites the entire commit history in your name.
- Alters commit timestamps to fall within the hackathon period using randomized intervals — simulating an "all-nighter grind."
How we built it
- Used Git + Python to rebase repos with new commit authors and timestamps.
- Queried GitHub’s API using LLama AI-generated keywords from the stolen codebase.
- Selected top 20 similar repos based on code structure.
- Compared the code for similar lines.
Accomplishments that we're proud of
- Successfully simulated realistic cheating behavior.
- Built a working detection pipeline to reverse-engineer and flag plagiarism.
- Tied both tools into a real AI safety and control use case.
What we learned
- AI can easily be used to manipulate trust-based systems.
- Red-teaming is critical for building guardrails.
- You can’t detect bad actors without first thinking like one.
What's next for Cheatathon
- Using AI to rewrite the codebase
- Add deeper analysis for commit graphs and contributor behavior to catch odd commits.
Built With
- bolt
- githubapi
- llama
- orchids
Log in or sign up for Devpost to join the conversation.