Dragonfly

Built an autonomous security testing engine with model-agnostic execution-time safeguards. Wraps 114 security testing tools behind a composable policy pipeline with scope enforcement, action gating, and kill switches. Provides adapters for MCP, CLI, and LangChain/LangGraph. Ran a scripted ablation study (45 runs) and an LLM behavioral study (60 runs) to evaluate safeguard effectiveness. Includes 1,259 tests and a Docker test range.