National Cyber Warfare Foundation (NCWF)login

Inside VulnBot: a multi-agent LLM framework that mimics human pentest workflows


0 user ratings
2026-11-10 09:29:01
milo
Red Team (CNA)
"Inside

VulnBot is a research-driven framework that uses LLM-based multi-agent collaboration, Kali Linux tooling and RAG knowledge retrieval to automate authorized penetration testing engagements.








ToolKHenryAegis/VulnBot — LLM multi-agent framework that replicates human penetration testing team workflows
CategoryAutonomous penetration testing / multi-agent LLM automation (Python)
Primary UseAutomating authorized security assessments by orchestrating Kali Linux tooling through collaborating LLM agents with RAG-backed knowledge retrieval
Safe UseIntended for authorized penetration tests, academic research (companion arXiv paper 2501.13411) and lab environments with explicit written permission from system owners
Telemetry NoteRequires a configured MySQL database and external Kali Linux host, so agent activity, task records and tool invocations are logged server-side and visible to defenders monitoring those systems

VulnBot, hosted at KHenryAegis/VulnBot under an MIT license, is an academic-minded automated penetration testing framework that applies Large Language Models to a problem that has resisted clean automation for years: reproducing the collaborative, judgment-heavy workflow of a human pentest team. Rather than a single scripted pipeline, VulnBot organizes LLM-driven agents into a multi-agent system where responsibilities are divided the way a real engagement would divide them, and the framework orchestrates Kali Linux tooling on behalf of those agents. The project ships with roughly 224 stars and is written in Python 3.11.11, and it is backed by a peer-style paper published as arXiv:2501.13411 — a signal that the authors are approaching this as a research contribution rather than a commercial product.


The core architectural idea is separation between reasoning and execution. The LLM agents plan, delegate and interpret results, while the actual offensive tooling lives on a separately configured Kali Linux host whose hostname, port, username and password are supplied in the configuration guide. This split matters for anyone studying the design: the agents never execute raw capability themselves, they dispatch work to a dedicated attack box, which is also the natural chokepoint where authorization boundaries and logging should be enforced. It is a sensible pattern for an authorized engagement, because scope enforcement can be applied at the execution layer rather than trusted entirely to model output.


The second major architectural pillar is retrieval-augmented generation. VulnBot integrates a RAG module — explicitly built on top of Langchain-Chatchat, which the README credits with a thank-you note — backed by a milvus vector store and a configurable knowledge base name (kb_name). Enabling it is a single toggle: set enable_rag to true and point it at the right milvus instance. Functionally, this gives the agents a grounded knowledge layer of security documentation and technique references instead of relying purely on what the base model memorized, which in theory reduces hallucinated tool flags and invented vulnerabilities during analysis.


Persistence and coordination run through a MySQL database whose host, port, user, password and database settings are part of the same configuration surface. For a multi-agent system this is the shared state store — the place where engagement context, discovered information and task progression live between agent interactions. From a defensive-analysis standpoint, that database is also a useful audit artifact: an assessor reviewing what the framework actually did can inspect the record of agent decisions and dispatched tooling rather than reverse-engineering behavior from shell history alone.


The LLM layer itself is deliberately provider-agnostic. Configuration exposes base_url, llm_model_name and api_key, meaning the framework talks to any OpenAI-compatible endpoint — a local model server, a hosted API, or an internal gateway. That flexibility is practically important for consulting teams with data-handling policies that prohibit sending client engagement data to third-party APIs, since a self-hosted model keeps the entire reasoning loop in-house. It also means the framework's behavior is only as good as the model backing it, which is a honest limitation worth stating.


Operationally, the lifecycle is compact. After git clone https://github.com/KHenryAegis/VulnBot and pip install -r requirements.txt, the user edits the configuration per the Configuration Guide.md document, then runs python cli.py init to initialize the project, python cli.py start -a to bring up the RAG module, and finally python cli.py vulnbot -m {max_interactions} to launch the agent loop with a bounded number of interactions. That -m cap is a thoughtful safety valve: it bounds how far the autonomous loop can run before a human reviews the results, which is exactly the kind of human-in-the-loop control that authorized assessments require.


What the README does not spell out — and what a careful reader should infer from the paper reference — is the exact division of labor among agents. The arXiv citation (2501.13411, Kong, Hu, Ge, Li, Li and Wu, 2025, primary class cs.SE) frames the work as software engineering research into multi-agent collaboration, so the interesting experimental question is how the agent roles, inter-agent communication and task decomposition compare to single-agent pentest automators. Anyone evaluating the framework for real use should read the paper alongside the code, since the README is intentionally thin on evaluation results and benchmark claims.


In terms of fit within an authorized workflow, VulnBot sits in the same emerging category as other LLM-driven offensive automation frameworks: it is best understood as a force multiplier for a licensed engagement, not a push-button attack tool. The configuration requirements alone make that clear — you need to control a Kali Linux host, a MySQL instance and an LLM endpoint before anything happens. Responsible use means explicit written scope, network isolation for the target range, and the interaction cap treated as a hard review boundary rather than an inconvenience.


For defenders and blue-team researchers, the telemetry profile is fairly distinctive. An engagement running VulnBot implies a milvus vector database, MySQL activity, and a stream of API calls to an LLM endpoint from the orchestrating machine, plus whatever footprint the underlying Kali tools leave on target systems. Purple teams can leverage that: instrument the range, watch how the agent loop sequences traditional tooling, and use the recorded agent decisions in MySQL to build detection narratives that map machine-driven activity patterns against human-operator baselines.


As a project, VulnBot is a credible research artifact with clean licensing, a published paper, and an architecture that takes grounding (RAG), bounded autonomy (max_interactions) and execution separation (dedicated Kali host) seriously. Its README is honest about what it borrows — the Langchain-Chatchat lineage for the RAG layer is credited directly — and issues and pull requests are explicitly welcomed. For professionals tracking where agentic LLM systems are taking offensive security automation, this repository is a worthwhile read, both for its engineering choices and for the defensive questions it raises about machine-paced assessment activity.


The bottom line for an authorized practitioner: VulnBot is a framework to study, replicate and evaluate in a lab before it touches anything client-facing. Its value today is educational and experimental — a working demonstration that human pentest team workflows can be decomposed into cooperating LLM agents with grounded knowledge retrieval — and its bounded-interaction design shows the authors understand that autonomy in security testing must remain reviewable. That framing, more than any single feature, is what makes this repository worth watching.



Official project repository for KHenryAegis/VulnBot.

Download Tool

Educational analysis for authorized security professionals. Use only in controlled, authorized environments.






Source: OffensiveSec
Source Link: https://www.offsecblog.com/2026/11/inside-vulnbot-multi-agent-llm.html


Comments
new comment
Nobody has commented yet. Will you be the first?
 
Forum
Red Team (CNA)



Copyright 2012 through 2026 - National Cyber Warfare Foundation - All rights reserved worldwide.