Your New Red-Teamer Is Three LLMs and a Whiteboard of Graphs

What it does

LuaN1aoAgent is an autonomous penetration-testing system that uses Large Language Models to move from reconnaissance through exploitation without human intervention. It breaks the work into three roles—Planner, Executor, and Reflector—that collaborate over an event bus, while a dynamic directed acyclic graph tracks task states and dependencies in real time. The README claims a 90.4% success rate on benchmark tasks with a median exploit cost of $0.09.

The interesting bit

Instead of letting a single LLM hallucinate attack chains, the agent builds explicit causal graphs where every hypothesis must be anchored to scanned evidence and quantified with confidence scores. When the Executor hits a dead end, the Reflector performs L1-L4 failure attribution so the Planner can prune or branch the task graph locally rather than regenerating the whole plan.

Key highlights

• Dual-graph architecture: a causal graph drives evidence-based decisions, while a separate DAG handles parallel task scheduling and dynamic replanning.
• P-E-R (Planner-Executor-Reflector) framework keeps roles decoupled; the Planner edits the task graph via structured operations like ADD_NODE and DEPRECATE_NODE rather than natural-language chatter.
• Built-in Human-in-the-Loop mode lets operators approve, reject, or inject tasks through a Web UI or CLI before high-risk actions run.
• Tool integration uses the Model Context Protocol (MCP) to orchestrate shell commands, HTTP requests, and Python execution; the README warns these are high-privilege and recommends containerized isolation.
• A standalone Web UI visualizes live graph evolution and persists task history in SQLite.

Caveats

• The project bundles high-privilege primitives like shell_exec and python_exec, so the authors explicitly recommend running it inside Docker or another isolated environment.
• Several roadmap items—Metasploit RPC integration, multimodal traffic analysis, and reinforcement-learning-based self-evolution—are still unchecked.

Verdict

Security researchers and red-teamers experimenting with LLM-driven automation should take a look; if you need a production-hardened scanner with fixed compliance reports, this is still a research-grade agent.