This model was trained 2x faster with
Unsloth and Huggingface's TRL library.
VulnHunter is a fine-tuned Qwen2.5-Coder-7B model specialized for security vulnerability detection and patching. It was trained using GRPO (Group Relative Policy Optimization) with a custom security reward function.
1from unsloth import FastLanguageModel
2
3model, tokenizer = FastLanguageModel.from_pretrained(
4 "gateremark/vulnhunter-agent"
5)
6
7# Analyze vulnerable code
8prompt = """Analyze this code for security vulnerabilities:
9query = f"SELECT * FROM users WHERE id = {user_id}"
10cursor.execute(query)
11"""
12
13inputs = tokenizer(prompt, return_tensors="pt")
14outputs = model.generate(**inputs, max_new_tokens=256)
15print(tokenizer.decode(outputs[0], skip_special_tokens=True))
GRPO eliminates the critic model by comparing responses within groups, giving PPO-quality learning without 2x memory overhead.
1# Input
2query = f"SELECT * FROM users WHERE username = '{username}'"
3
4# VulnHunter Output
5# "SQL injection vulnerability. Use parameterized queries:
6# query = 'SELECT * FROM users WHERE username = %s'
7# cursor.execute(query, (username,))"
1# Input
2return f"<h1>Results for: {query}</h1>"
3
4# VulnHunter Output
5# "XSS vulnerability. Escape user input:
6# from html import escape
7# return f'<h1>Results for: {escape(query)}</h1>'"
1# Input
2return send_file(f"/uploads/{filename}")
3
4# VulnHunter Output
5# "Path traversal vulnerability. Use safe_join:
6# from werkzeug.utils import safe_join
7# return send_file(safe_join('/uploads', filename))"
1git clone https://github.com/gateremark/vulnhunter
2cd vulnhunter
3uvicorn vulnhunter.env_server.server:app --port 8000
1cd vulnhunter/green_agent
2python server.py
3# Agent at http://localhost:9009
1{
2 "name": "VulnHunter",
3 "skills": [{"id": "analyze_code", "name": "Analyze Code"}]
4}
1@misc{vulnhunter2026,
2 author = {gateremark},
3 title = {VulnHunter: AI Security Agent with GRPO},
4 year = {2026},
5 publisher = {HuggingFace},
6 url = {https://huggingface.co/gateremark/vulnhunter-agent}
7}
