msoedov/agentic_security

What it solves

Agentic Security is an open-source vulnerability scanner designed to protect Large Language Models (LLMs) and agent workflows from security threats. It helps developers and security teams proactively identify weaknesses such as jailbreaks, fuzzing vulnerabilities, and multimodal attacks before they are deployed into production.

How it works

The tool operates by probing LLM APIs using a variety of attack vectors. It uses a plain-text HTTP specification to connect to target models and replaces placeholders with specific attack prompts. It can load datasets from Hugging Face, local CSV files, or generate dynamic datasets using mutations (like Base64 encoding or ROT13) to bypass safety filters. The system supports text, image, and audio modalities and can be integrated into CI/CD pipelines as an automated security check.

Who it’s for

This project is intended for AI developers, security researchers, and security teams who need to stress-test their AI systems and ensure they are robust against adversarial attacks.

Highlights

• Multimodal Probing: Supports scanning for vulnerabilities across text, image, and audio inputs.
• Multi-Step Jailbreaks: Simulates iterative attack sequences to uncover deep safety weaknesses.
• RL-Based Attacks: Uses reinforcement learning to create adaptive probes that evolve against model defenses.
• CI/CD Integration: Includes a sample GitHub Action for automated security scanning during development.
• Extensible Datasets: Allows users to add custom CSV datasets or integrate existing datasets from Hugging Face.