jackysnake/RedCoder
RedCoder by jackysnake is an 8 billion parameter model fine-tuned for automated multi-turn red teaming of Code LLMs. It specializes in generating adversarial prompts to induce CWE vulnerabilities like path traversal and SQL injection in target models. This model is designed for security benchmarking, red-teaming evaluations, and research into adversarial prompt generation for secure LLM development.
Loading preview...
REDCODER: Automated Multi-Turn Red Teaming for Code LLMs
REDCODER is an 8 billion parameter language model developed by jackysnake, specifically fine-tuned for adversarial multi-turn prompt generation. Its primary function is to engage target Code LLMs in conversations that gradually lead them to generate code containing CWE vulnerabilities, such as path traversal or SQL injection.
Key Capabilities
- Adversarial Prompt Generation: Creates multi-turn prompts designed to exploit vulnerabilities in Code LLMs.
- Vulnerability Induction: Focuses on inducing specific CWE vulnerabilities in generated code.
- Research Tool: Intended for academic and research settings to study LLM security.
Intended Use Cases
- Red-teaming evaluations for Code LLMs to identify weaknesses.
- Security benchmarking of model guardrails and filters.
- Multi-turn adversarial prompt generation in research contexts.
Important Note
This model is explicitly not intended for generating real-world exploits. Its use is strictly limited to research, safety evaluation, and the development of more secure LLMs. Further details and the underlying research can be found in the associated paper: [arXiv:2507.22063].