SecureCode: Comprehensive Security Training Dataset for AI Coding Assistants

Examples Languages CWEs Frameworks

The largest open security training dataset for AI coding assistants, covering both traditional web security and AI/ML security

Overview

Quick Start

from datasets import load_dataset
# Load everything (2,372 examples)
dataset = load_dataset("scthornton/securecode")
# Load only web security examples (1,625)
web = load_dataset("scthornton/securecode", "web")
# Load only AI/ML security examples (747)
aiml = load_dataset("scthornton/securecode", "aiml")

Dataset Configs

SecureCode Dataset Family

Web Security Coverage (1,625 examples)

OWASP Top 10 2021 Categories

Languages and Frameworks

AI/ML Security Coverage (747 examples)

OWASP LLM Top 10 2025 Categories

Frameworks

Unified Schema

{
 "id": "example-id",
 "metadata": {
 "category": "Category name",
 "severity": "CRITICAL",
 "cwe": "CWE-79",
 "lang": "python"
 },
 "context": {
 "description": "Vulnerability description",
 "impact": "Business and technical impact"
 },
 "conversations": [
 {"role": "human", "content": "Developer asks about building a feature"},
 {"role": "assistant", "content": "Vulnerable code + secure code + defense-in-depth"},
 {"role": "human", "content": "Follow-up about testing and edge cases"},
 {"role": "assistant", "content": "Testing guidance + monitoring + common mistakes"}
 ],
 "validation": { ... },
 "quality_score": null,
 "security_assertions": [],
 "references": []
}

Quality

Usage Examples

Filter by OWASP Category

# Web: filter by OWASP Top 10 category (web category values are lowercase snake_case;
# filter on owasp_2021 for a stable match)
web = load_dataset("scthornton/securecode", "web")
injection = web["train"].filter(
 lambda x: x.get("metadata", {}).get("owasp_2021", "").startswith("A03")
)
# AI/ML: filter by LLM Top 10 category
aiml = load_dataset("scthornton/securecode", "aiml")
prompt_injection = aiml["train"].filter(
 lambda x: x.get("metadata", {}).get("owasp_llm_2025") == "LLM01"
)

Extract Training Pairs

dataset = load_dataset("scthornton/securecode")
for example in dataset["train"]:
 conversations = example["conversations"]
 for turn in conversations:
 if turn["role"] == "human":
 prompt = turn["content"]
 else:
 response = turn["content"]

Fine-Tuning with Transformers

from datasets import load_dataset
from transformers import AutoTokenizer
dataset = load_dataset("scthornton/securecode")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-3B")
def format_for_training(example):
 messages = [
 {"role": t["role"].replace("human", "user"), "content": t["content"]}
 for t in example["conversations"]
 ]
 return {"text": tokenizer.apply_chat_template(messages, tokenize=False)}
formatted = dataset["train"].map(format_for_training)

Citation

@misc{thornton2025securecode,
 title={SecureCode: A Production-Grade Multi-Turn Dataset for Training Security-Aware Code Generation Models},
 author={Thornton, Scott},
 year={2026},
 publisher={perfecXion.ai},
 url={https://huggingface.co/datasets/scthornton/securecode},
 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}
}

Ethics and Intended Use

License

Links