An LLM-driven system that autonomously red-teams networks: it takes a high-level attack strategy, decomposes it into concrete actions across multiple hosts, executes them, and adapts based on what it finds, with the resulting attack data used to train stronger automated defenders.