PurpCode UIUC

Welcome to the PurpCode project!

PurpCode is an alignment method and a fully open-source recipe (data, model, and code) for eliciting cybersafe reasoning capabilities of coding models, including secure code generation and defending against malicious cyber events. PurpCode includes two alignment stages:

1. Rule Learning: teaching LLMs secure coding rules and general safety practices
2. Reinforcement Learning: letting LLMs co-exercise their safety and utility via verifiable tasks

We also curate comprehensive safety data via internal red teaming and use various evaluators covering cybersafety, utility, and overrefusal.

• Paper: https://arxiv.org/abs/2507.19060
• GitHub: https://github.com/purpcode-uiuc/purpcode
• About: We are the 🥇Winner Defender Team at Amazon Nova AI Challenge 2025, from University of Illinois Urbana-Champaign!

To cite our work:

@article{purpcode,
  title = {PurpCode: Reasoning for Safer Code Generation},
  author = {Liu, Jiawei and Diwan, Nirav and Wang, Zhe and Zhai, Haoyu and Zhou, Xiaona and Nguyen, Kiet A. and Yu, Tianjiao and Wahed, Muntasir and Deng, Yinlin and Benkraouda, Hadjer and Wei, Yuxiang and Zhang, Lingming and Lourentzou, Ismini and Wang, Gang},
  journal = {arXiv preprint arXiv:2507.19060},
  year = {2025},
}