I've been experimenting with LangGraph's ReAct agents for offensive security automation and wanted to share some interesting results. I built an autonomous exploitation framework that uses a tiny open-source model (Qwen3:1.7b) to chain together reconnaissance, vulnerability analysis, and exploit execution—entirely locally without any paid APIs