The results from OpenAI’s gpt-oss-20b Red-Teaming Challenge on Kaggle are in, and my scores were poor. I won’t pretend this doesn’t sting. The findings themselves held up when I tested them, but the submission failed on technical errors, not on the implementation, which is a frustrating way to lose. It’s important to remember that a competition like this scores what the judges can reproduce, not what you know works, and that is a fair rule even when it costs me.
If nothing else, I learned more about the model’s failure modes in three weeks than I would have from reading about them. Next time, the submission gets as much testing as the attacks do.