Red‑Teaming Challenge - OpenAI gpt-oss-20b

Overview

This hackathon challenges you to find undiscovered vulnerabilities and harmful behaviors in OpenAI's gpt-oss-20b open weight model. You will submit up to five distinct issues with reproducible reports detailing your findings. The goal is to improve AI safety and shape the future of alignment tools for the open source community.

Requirements

Teams must submit up to five distinct issues, each with a reproducible report. Submissions require a Kaggle Writeup detailing the strategy, discovery process, tooling, threat analysis, and lessons learned. Findings should be uploaded as separate Kaggle Datasets in JSON format. An optional reproduction notebook and open-source tooling (notebook, package, or script directory) are encouraged and can improve scores. All findings and datasets should remain private until the competition deadline.

Prizes

There is a total prize pool of $500,000 across 10 awards. Each of the 10 winning submissions will receive $50,000. The competition does not award Kaggle points or medals.