AIPG: The AI Security Enclave

The AI Proving Grounds (AIPG) at WWT is an ensemble of cutting-edge hardware, software, and engineering expertise dedicate to all things AI. Clients across all industries and of all sizes are using the AI Proving Ground to accelerate, secure and scale their AI journeys. The AI Security Enclave is a small environment within the AIPG dedicated to supporting AI Security efforts, a place to explore the complicated and dynamic landscape of AI safely and with access to the wide range of expensive resources provided by the AIPG. Working with WWT engineers, clients can quickly design and implement test plans to gauge the security of AI applications as well as of the specialized tools providing security for AI.

• Provide access to a wide range of LLMs, either local or remote via API
• Act as a "middleware" capability to exercise ideas for deeper follow-on efforts like Proofs-of-Concept or Lab-as-a-Service
• Thorough hands-on training in programatically using, testing, and attacking AI applications and AI security solutions
• Allow rapid design and implementation for testing specific OEM solution
• Facilitate a wide range of projects in AI security, such as:
  • Testing efficacy of LLM firewalls 
  • Building malicious model files for testing model file scanning
  • Profiling execution times and latency of LLM input and output scanners
  • Functionality testing of integrating AI security solutions with AI web applications

Hardware

• Servers
  • Dell PowerEdge 640 (788 GB RAM) 
  • Dell PowerEdge 740 (788 GB RAM)
• Liqid mapped resources 
  • One Nvidia A100 GPU per server 
  • Multiple 800GB SSD drives
• Operating System: Ubuntu 22.04

Software

• Ollama (CLI and API access to local LLMs)
• OpenWebUI (high-level application testing)
• Python access to LLM APIs (e.g. OpenAI and Azure AI Service)
• Development environment: conda, python, vscode, jupyter notebooks, other Linux tools
• Open-source security packages for testing, including:
  • llm-guard (Protect AI)
  • NeMo Guardrails, Morpheus (Nvidia)
  • Purple Llama suite (Meta)
• WWT-developed test software and attack data 
  • Corpus of prompts used for jailbreaking, prompt injection
  • Python "test harness" used as template for new attack trials