https://github.com/jacktheprogrammer/sentinel-ml-security-crag-agent
A CRAG agent crafted using Gemini, langgraph and python to audit the ML scripts before deploying in production level
https://github.com/jacktheprogrammer/sentinel-ml-security-crag-agent
Last synced: 13 days ago
JSON representation
A CRAG agent crafted using Gemini, langgraph and python to audit the ML scripts before deploying in production level
- Host: GitHub
- URL: https://github.com/jacktheprogrammer/sentinel-ml-security-crag-agent
- Owner: JackTheProgrammer
- License: agpl-3.0
- Created: 2026-03-05T02:46:36.000Z (5 months ago)
- Default Branch: main
- Last Pushed: 2026-03-18T18:04:32.000Z (4 months ago)
- Last Synced: 2026-03-19T07:41:59.127Z (4 months ago)
- Language: Jupyter Notebook
- Size: 3.35 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 🛡️ Sentinel-ML: Security-First Agentic CRAG
Autonomous AI Agent System for Auditing ML Workflows via LangGraph & Corrective RAG.
## 🔍 Overview
**Sentinel-ML** is an intelligent security layer for the ML lifecycle. It doesn't just scan code; it understands the context of training, deployment, and inference scripts to intercept vulnerabilities before they reach production.
## 🚀 Why Sentinel-ML?
Standard LLMs often provide insecure code snippets or hallucinate outdated security patches. Sentinel-ML solves this by treating security as a Stateful Engineering Problem, not just a prompting task.
### 🧠 Architecture & Logic
Sentinel-ML is not a simple wrapper. It uses a **Directed Acyclic Graph (DAG)** via LangGraph to manage the audit state:
1. **Contextual Retrieval**: Proactively fetches framework-specific security signatures.
2. **Corrective Grading**: Self-evaluates retrieved data quality (`GOOD`, `AMBIGUOUS`, `INCORRECT`).
3. **Recursive Re-writing**: Dynamically optimizes queries to trigger targeted web searches when local knowledge is insufficient.
4. **Final Security Audit**: The final reasoning gatekeeper against RCE, Model Poisoning, and Path Injection.
## 📊 Agent flow
### Node sequence
### Internal execution flow
## ⚡Demonstration
**Problem**: Insecure model loading using `pickle` (Remote Code Execution risk).
**Sentinel-ML Solution**: Seamless migration to high-performance, safe alternatives like `safetensors`.
```python
res = app.invoke({
"question": "Find vulnerabilities in this model loading script and provide a secure, high-performance alternative.",
"vulnerable_code": """import pickle
import pandas as pd
def load_user_model(user_id):
# DANGEROUS: Loading model from untrusted user path
path = f"models/{user_id}_model.pkl"
with open(path, 'rb') as f:
model = pickle.load(f) # <--- Potential RCE Vulnerability
return model
""",
"query": "Find vulnerabilities in this model loading script and provide a secure, high-performance alternative.",
"web_docs": [],
"fixed_code_cost_reduced": "",
"entire_fixed_code": "",
"final_report": None,
"display_report": "",
"docs": [],
"good_docs": [],
"kept_strips": [],
"strips": [],
"refined_context": "",
"recommend_fix_only": False,
"verdict": " ",
"reason": " ",
"web_queries": [],
"fix_recommendations": set()
})
print("Fixed code with reduced computational cost:\n", res['fixed_code_cost_reduced'])
```
### Output
```text
Fixed code with reduced computational cost:
from pathlib import Path
from safetensors.torch import load_file
def load_user_model(user_id):
# Optimized safe loading: uses zero-copy memory mapping via safetensors
# Minimalist path sanitization to reduce overhead
safe_id = Path(user_id).name
path = Path("models") / f"{safe_id}_model.safetensors"
# load_file is significantly faster than pickle and inherently safe
return load_file(str(path))
```
## 🛠️ Key Capabilities
- **Vulnerability Detection**: Identifies Data Leakage, Model Poisoning, and Adversarial Vectors.
- **Best Practices Recommendations**: Uses CRAG to verify all recommendations against live web data.
- **Automated Analysis**: Provides deep-dive security reports with zero manual intervention.
- **Integration with CI/CD**: State-persistent logic designed for automated pipeline security gates.
## 📦 Installation and setup
To use the Sentinel-ML Security CRAG AI-agent, follow these steps:
1. Clone the repository:
```bash
git clone https://github.com/JackTheProgrammer/Sentinel-ML-security-CRAG-agent.git
```
2. Install the required dependencies:
```bash
pip install -r configs/requirements.txt
```
3. Use the code from the jupyter notebook `notebooks/sentinel_ml_security_crag_demo.ipynb` to see the agent in action.