REDCODER: Automated Multi-Turn Red Teaming for Code LLMs

🔬 A model fine-tuned for adversarial multi-turn prompt generation to induce vulnerabilities in Code LLMs.
📄 [arXiv:2507.22063] • 🧠 💻 Full code & data: GitHub – luka-group/RedCoder

---

🧠 Model Summary

REDCODER is a red-teaming LLM trained to engage target Code LLMs in multi-turn conversations that gradually steer them into generating CWE vulnerabilities (e.g., Such as path traversal, SQL injection, etc.).

This model is designed to support:

• ⚔️ Red-teaming evaluations for Code LLMs
• 🧪 Security benchmarking of model guardrails and filters
• 🧩 Multi-turn adversarial prompt generation in research settings

⚠️ This model should not be used to generate real-world exploits. Its intended use is for research, safety evaluation, and secure LLM development.

---

If you find this work useful, please cite:

@article{mo2025redcoder,
  title   = {REDCODER: Automated Multi-Turn Red Teaming for Code LLMs},
  author  = {Wenjie Jacky Mo and Qin Liu and Xiaofei Wen and Dongwon Jung and
             Hadi Askari and Wenxuan Zhou and Zhe Zhao and Muhao Chen},
  journal = {arXiv preprint arXiv:2507.22063},
  year    = {2025}
}