SecureCode

Comprehensive security training dataset for AI coding assistants — 2,185 examples covering both traditional web security and AI/ML security.

Built by perfecXion.ai.

Dataset Family

Dataset	Examples	Focus	HuggingFace	GitHub
SecureCode	2,185	Unified (web + AI/ML)	scthornton/securecode	This repo
SecureCode v2	1,435	Web security (OWASP Top 10 2021)	scthornton/securecode-v2	securecode-v2
SecureCode AI/ML	750	AI/ML security (OWASP LLM Top 10 2025)	scthornton/securecode-aiml	securecode-aiml

Quick Start

from datasets import load_dataset

# Load everything (2,185 examples)
dataset = load_dataset("scthornton/securecode")

# Load only web security (1,435 examples)
web = load_dataset("scthornton/securecode", "web")

# Load only AI/ML security (750 examples)
aiml = load_dataset("scthornton/securecode", "aiml")

What's In It

Every example is a 4-turn conversation between a developer and an AI coding assistant. The developer asks how to build something, and the assistant provides a vulnerable implementation, explains why it's dangerous, shows a secure alternative with 5+ defense layers, and then covers testing, monitoring, and common mistakes.

Web Security (1,435 examples): SQL injection, XSS, authentication bypass, SSRF, cryptographic failures, and more across 12 programming languages and 9 web frameworks (Express.js, Django, Spring Boot, Flask, Rails, Laravel, ASP.NET Core, FastAPI, NestJS).

AI/ML Security (750 examples): Prompt injection, model poisoning, embedding manipulation, system prompt leakage, excessive agent autonomy, and more across 30+ AI/ML frameworks (LangChain, OpenAI, Anthropic, HuggingFace, LlamaIndex, ChromaDB, vLLM, CrewAI, AutoGen, etc.).

Unified Schema

All conversations use a normalized {role, content} format:

{
  "id": "example-id",
  "metadata": { "category": "...", "severity": "CRITICAL", "cwe": "CWE-79", "lang": "python" },
  "context": { "description": "...", "impact": "..." },
  "conversations": [
    {"role": "human", "content": "How do I build secure JWT auth?"},
    {"role": "assistant", "content": "Here's the vulnerable version... here's the secure version..."},
    {"role": "human", "content": "How do I test this?"},
    {"role": "assistant", "content": "Here's how to test, monitor, and avoid common mistakes..."}
  ],
  "quality_score": null,
  "security_assertions": [],
  "references": []
}

Building the Unified Dataset

The unified dataset is built from the two source datasets using a normalization script that converts v2.x conversations from {turn, from, value} to {role, content} format.

python3 scripts/build_unified_dataset.py

This generates unified-data/data/web/ (1,435 files) and unified-data/data/aiml/ (750 files), ready to push to HuggingFace.

Configs

Config	Examples	OWASP Standard
default	2,185	Both
web	1,435	OWASP Top 10 2021
aiml	750	OWASP LLM Top 10 2025

Citation

@misc{thornton2025securecode,
  title={SecureCode v2.0: A Production-Grade Dataset for Training Security-Aware Code Generation Models},
  author={Thornton, Scott},
  year={2025},
  publisher={perfecXion.ai},
  url={https://huggingface.co/datasets/scthornton/securecode-v2},
  note={arXiv:2512.18542}
}

@dataset{thornton2026securecodeaiml,
  title={SecureCode AI/ML: AI/ML Security Training Dataset for the OWASP LLM Top 10 2025},
  author={Thornton, Scott},
  year={2026},
  publisher={perfecXion.ai},
  url={https://huggingface.co/datasets/scthornton/securecode-aiml}
}

License

• Web examples: CC BY-NC-SA 4.0
• AI/ML examples: MIT
• Unified dataset: CC BY-NC-SA 4.0 (the more restrictive of the two)

---

Contact

Scott Thornton — AI Security Researcher

• Website: perfecxion.ai
• Email: [email protected]
• LinkedIn: linkedin.com/in/scthornton
• ORCID: 0009-0008-0491-0032
• GitHub: @scthornton

Security Issues: Please report via SECURITY.md