Agentic Radar: Visualization Tool for Agentic Workflow Security Inspection

General Introduction

Agentic Radar is an open source tool from SplxAI designed for scanning and securing large model (LLM)-based AI workflows. It analyzes code and generates clear workflow diagrams and security reports to help developers and security experts quickly understand how AI systems work and identify potential vulnerabilities. Tool Support CrewAI cap (a poem) LangGraph The test results can also be checked against security standards such as OWASP LLM Top 10, making it easy for users to locate problems. It is easy to use, and all operations are done locally to ensure code security, which is suitable for improving the transparency and security of AI projects.

Function List

• Scanning AI workflows: Analyze the code to sort out the structure and operational logic of the AI system.
• Generate workflow diagrams: Draw visual dependency diagrams showing the relationships between modules and tools.
• List of identification tools: Lists all external and customized tools used by the system.
• Detecting Security Vulnerabilities: Identify potential risks and match criteria such as OWASP LLM Top 10.
• Support for multiple frameworks: Compatible with CrewAI and LangGraph, with plans to support more frameworks in the future.
• Output detailed reports: Generate security reports in HTML format for easy viewing and sharing.
• local operation: All scanning is done on the user's device and no code is uploaded.

Using Help

Agentic Radar is a command line tool that is easy to install and use. Below are detailed steps to help you get started from scratch.

Installation process

1. Preparing the environment
   • Requires Python 3.9 or above, run python --version Check.
   • Install Git for downloading code from GitHub.
   • Install graphviz and cairo for generating workflow diagrams:
     • Mac: Using Homebrew, run brew install graphviz cap (a poem) brew install cairoThe
     • Linux: Run sudo apt-get install graphviz cap (a poem) sudo apt-get install libcairo2The
     • Windows (computer): Install with Chocolatey, run choco install graphviz cap (a poem) choco install cairo, or manually download the installer.
2. Installing Agentic Radar
   • Open a terminal and clone the code:

     git clone https://github.com/splx-ai/agentic-radar.git
     

   • Go to the catalog:

     cd agentic-radar
     

   • Installation Tools:

     pip install .
     

   • Check the version to confirm successful installation:

     agentic-radar --version
     

     For example, returning 0.2.0, that means there's no problem.

3. Addressing dependencies
   • If an error is reported about missing libraries, install them manually:

     pip install pydot pydantic typer-slim jinja2
     

Usage

Agentic Radar is easy to use by scanning code and generating reports from the command line.

Basic Scan

• Runs in the terminal:

  agentic-radar -i . /my_project -o report.html

• -i Specify the code folder, e.g. . /my_projectThe
• -o Specify the output report file name, for example report.htmlThe

Specify the framework

• If using CrewAI:

  agentic-radar -i . /my_project -o report.html crewai
  

• If you use LangGraph:

  agentic-radar -i . /my_project -o report.html langgraph
  

View Report

• After running, open the report.html, included inside:
  • workflow diagram: Shows module relationships with nodes and arrows that can be zoomed in and out for viewing.
  • List of tools: Lists all the tools used, such as APIs or custom functions.
  • vulnerability table: Label the issue and correspond to the OWASP LLM Top 10, e.g. "LLM02: Data Leakage".
  • suggestion: Give fixes such as adding input validation.

sample operation (computing)

Let's say you have a CrewAI program in . /chatbot Folder:

1. Running:

   agentic-radar -i . /chatbot -o chatbot_report.html crewai
   

2. show (a ticket) chatbot_report.htmlSee:
   • How to call the search tool for chat agents.
   • If the search tool is not encrypted, the report will warn.
   • It is recommended that you add API key protection.

Featured Function Operation

1. Use of workflow diagrams
   • Each node in the graph is a module and the arrows are dependencies. For example, "Data Processing" points to "Output Generation".
   • Drag or zoom with your mouse to view details.
2. vulnerability testing
   • The report will list the problem, such as "prompt injection" (LLM01), and indicate the line of code.
   • Fixed as suggested, e.g. limiting the length of user input.
3. Extended Support
   • If you want to support the new framework, you can change the code and submit it to GitHub to participate in community development.

Frequently Asked Questions

• The command is invalid.: Check Python and PATH settings.
• Figure not shown: Make sure graphviz and cairo are installed and restart the terminal.
• Less reporting: Check the input folder to make sure you have the full code.

Getting Help

• (of a computer) run agentic-radar --help See more options, such as adjusting the output path.

application scenario

1. Debugging complex projects
   The AI system has too many modules and is running out of order. Sweep it with Agentic Radar, and you can see in the picture which parts are not connected.
2. Pre-launch inspection
   Projects are going live, use it to scan and make sure there are no vulnerabilities, especially where user data is concerned.
3. Teamwork
   When developing with multiple people, the generated workflow diagrams allow everyone to quickly understand the code structure and improve efficiency.
4. Learning framework
   When new to CrewAI, run the tool and look at the diagrams to understand how the framework organizes the code.

QA

1. Does Agentic Radar cost anything?
   Totally free, open source project, code on GitHub at will.
2. What languages are supported?
   Only Python is supported now, since AI frameworks are mostly written in Python.
3. Will the code be uploaded?
   No, the whole thing runs locally and the code doesn't leave your device.
4. How do I update it?
   Go to GitHub, pull the latest code, and run it again. pip install .The
5. Will other frameworks be supported in the future?
   Will, plan to add LlamaIndex, Swarm, AutoGen, etc., depending on community contributions.