22.1.1 Hardware and software requirements

Hardware Specifications

Your primary machine, the “host,” will run virtualization software to create isolated “guest” machines for your attack tools and target systems. This host/guest architecture demands sufficient resources to avoid performance bottlenecks. While you can start with modest hardware, investing in a capable machine will significantly improve your workflow, especially when dealing with large language models (LLMs) or complex simulations.

The following table provides a tiered guide to hardware specifications. Aim for the “Recommended” tier for a smooth experience, but the “Minimum” will suffice for basic prompt injection and API-level testing.

The GPU Consideration

While not mandatory for all AI red teaming tasks, a dedicated GPU is a powerful accelerator. Many adversarial attacks, particularly those involving model fine-tuning, gradient-based optimization, or running local open-source models for analysis, are computationally infeasible without one.

For practical purposes, focus on NVIDIA GPUs due to the widespread adoption of the CUDA parallel computing platform in the machine learning ecosystem. The amount of Video RAM (VRAM) is often more critical than raw processing power, as it determines the size of the models you can load into memory. An 8GB VRAM card is a functional starting point, but 16GB or more will open up a much wider range of models and techniques.

Core Software Stack

With your hardware in place, the next layer is the software that will manage your environment. The key principle here is isolation: your red teaming activities must be sandboxed from your primary operating system and network.

Host Operating System

Your host OS should be stable, secure, and well-supported. A Linux distribution is highly recommended due to its robust virtualization capabilities, powerful command-line interface, and native support for many security tools.

• Recommended: Ubuntu 22.04 LTS or newer, Fedora Workstation. These offer a balance of modern packages and long-term stability.
• Viable Alternatives: macOS (with its Unix-like foundation) or Windows 10/11 using the Windows Subsystem for Linux (WSL2) for a near-native Linux experience.

Virtualization Software (Hypervisor)

The hypervisor is the most critical piece of software for creating your lab. It allows you to run multiple, isolated operating systems (virtual machines) on a single physical machine. This is non-negotiable for security and reproducibility.

• Oracle VirtualBox: Free, open-source, and cross-platform. An excellent choice for beginners.
• VMware Workstation (Windows/Linux) / Fusion (macOS): Commercial products known for high performance and advanced features like VM snapshots and cloning.
• KVM/QEMU: The native Linux virtualization solution. Extremely powerful and efficient, but managed primarily through the command line or with tools like virt-manager.

Quick System Check

Before proceeding, you can quickly verify your system’s core specifications from the command line on a Linux-based system. These commands provide a snapshot of your available resources.

# Check CPU information (cores, architecture)
lscpu | grep "Core(s) per socket"

# Check available RAM
free -h

# Check available disk space on mounted filesystems
df -h

# If you have an NVIDIA GPU and drivers installed
nvidia-smi

With these foundational hardware and software components selected, you are ready to move on to configuring your virtual environment, which is the next critical step in building a secure and functional AI red teaming lab.