AIGoat : A Deliberately Vulnerable AI Infrastructure

AI-Goat is an innovative open-source platform designed to address the growing need for hands-on training in AI security.

Developed by Orca Security, it provides a deliberately vulnerable AI infrastructure hosted on AWS, simulating real-world environments to highlight security risks associated with machine learning (ML) systems.

By focusing on the OWASP Machine Learning Security Top 10 risks, AI-Goat equips security professionals and researchers with practical tools to identify and mitigate vulnerabilities in AI applications.

Core Features And Objectives

AI-Goat aims to educate users about the intricacies of AI security through realistic scenarios. Its primary objectives include:

• AI Security Testing and Red-Teaming: Users can explore vulnerabilities in ML models and infrastructure.
• Infrastructure as Code (IaC): Leveraging Terraform and GitHub Actions, the deployment process is streamlined, offering a modular approach to learning.
• Risk Identification: It emphasizes understanding risks across AI applications, including data poisoning, supply chain attacks, and output integrity issues.

The infrastructure is structured into modules, each representing distinct AI applications with varying tech stacks such as AWS, React, Python 3, and Terraform.

AI-Goat incorporates three key challenges based on OWASP ML Security Top 10 risks:

1. AI Supply Chain Attack: Exploits vulnerabilities in the product search module by compromising the supply chain through malicious file uploads.
2. Data Poisoning Attack: Demonstrates how attackers can manipulate training datasets to alter personalized product recommendations.
3. Output Integrity Attack: Highlights weaknesses in content filtering systems, allowing users to bypass restrictions.

Deployment is simplified through Terraform workflows. Users can fork the repository, configure AWS credentials via GitHub secrets, and execute the deployment process. Manual installation is also supported for advanced users.

AI-Goat is ideal for:

• Security professionals seeking hands-on experience in AI risk mitigation.
• Organizations aiming to enhance their defenses against AI-specific threats.
• Researchers exploring vulnerabilities in ML systems.

By providing a controlled environment for experimentation, AI-Goat fosters a deeper understanding of potential threats while promoting best practices in securing AI infrastructures.