CrowdStrike Research: Securing AI-Generated Code with Multiple Self-Learning AI Agents

This document outlines CrowdStrike’s research into a multi-agent AI workflow that integrates LLMs with SAST and automated red-teaming to identify, validate (via exploitation), and patch code vulnerabilities, claiming major efficiency gains through continuous learning and collaboration among specialized agents. It describes how Vulnerability, Patching, and Red Teaming agents use README/context analysis, custom SAST rules, test generation, and RAG to refine detection accuracy and remediation, and positions the work as thought leadership showcased at NVIDIA GTC 2025.