VulnAgent

VulnAgent is an experimental AI-assisted cybersecurity agent framework that orchestrates modular agents using LLMs and domain-specific tools, communicating over XMPP. It combines reasoning with custom vulnerability tooling (e.g., severity and CWE classification, Vulnerability-Lookup API) to automate tasks such as vulnerability classification and interaction with security workflows.

While the concept of AI agents—models coupled with tools and orchestration logic—has become fairly standardized, VulnAgent explores a distinctive approach tailored to cybersecurity. Its agents communicate over XMPP, leveraging native support for asynchronous messaging, concurrent behaviours, presence, and discovery, making it well suited for distributed, agentic security workflows.

Features

• Modular AI agents combining reasoning (LLM) and tools
• Tool orchestration with clear mental models
• XMPP-based communication between agents
• Integration with the Vulnerability-Lookup API and custom classifiers (e.g., CWE and severity classification)

Architecture

Inter-agent communication

graph LR
    Ch[Chat Agent] <--> A[LLMAgent]
    A --> C[ContextManager]
    A --> D[LLMProvider]
    A --> E[LLMTool]
    D --> F[OpenAI/Ollama/etc]
    E --> I[Human-in-the-Loop]
    E --> T1[VLAI Severity - Text Classification]
    E --> T2[VLAI CWE - Text Classification]
    E --> T3[Vulnerability-Lookup API]
    E --> J[MCP]
    J --> K[STDIO]
    J --> L[HTTP Streaming]

Loading

Human-in-the-loop is still in work and will be probably linked to the Vulnerability-Lookup API tool.
The LLM provider can be configured in vulnagent.agent.llm:get_llm_provider(). The default is qwen2.5:7b.

Component Overview:

Component	Description
ChatAgent	Entry point optionnaly with guardrails filtering.
LLMAgent	Core agent that reasons using a language model.
ContextManager	Tracks conversation state and memory.
LLMProvider	Connects to models (OpenAI, Ollama, Qwen, etc.).
LLMTool	Performs actions such as classification, API queries, or human-in-the-loop checks.
MCP	Multi-channel publisher for STDIO or HTTP streaming outputs.

The LLMAgent (Qwen) leverages the VLAI Severity classification and VLAI CWE classification models as integrated tools, enabling automated vulnerability severity assessment and CWE categorization within its reasoning workflow.

Agent Principle

VulnAgent
 ├── Reasoning (LLM via spade-llm, Ollama or API)
 ├── Tools
 │    ├── SeverityClassifierTool (RoBERTa)
 │    ├── CVSS normalizer tool (planned)
 │    └── Other extensible tools
 └── Actions / Messages

You: "What is the severity of the vulnerability described ..."
LLM: "This looks like a vulnerability description.
      I should classify severity."
→ calls severity_classifier tool
→ receives result
→ explains or forwards

Tools are assigned to an (LLM) agent. An agent can use one or multiple tools and should clearly explain their functionality. Communications via XMPP/FIPA.

Test

Install Ollama

curl -fsSL https://ollama.ai/install.sh | sh
ollama pull llama3.1:8b
ollama pull qwen2.5:7b
ollama serve

# Check if default ports are already in use
netstat -an | grep 5222

# Try different ports if needed, or shutdown prosodyctl
spade run --client_port 6222 --server_port 6269

then use the Web interface to create the agent's password.

Alternatively (maybe even better, and it's what had been tested so far), use Prosody. In this case create the agent's password:

$ sudo prosodyctl adduser tool_assistant@localhost
$ sudo prosodyctl adduser user@localhost
$ sudo prosodyctl adduser coordinator@localhost

Install the project

$ cd VulnAgent/
$ poetry install
$ poetry shell

Launch the agents

$ vulnagent-llm
Device set to use cpu
XMPP server domain (default: localhost): 
LLM provider to use (default: qwen2.5:7b):   
Agent name (default: tool_assistant): 
LLM agent password: 
LLM Agent Web Interface: http://127.0.0.1:10000/spade
Press Ctrl+C to exit.

$ vulnagent-chat 
XMPP server domain (default: localhost): 
Agent name (default: chat_agent): 
Chat agent password: 
✅ Agent started!
🔧 Available tools:
• classify_severity
• classify_cwe
• get_current_time
• calculate_math
• get_weather

💡 Try these queries:
• 'What's the severity of the vulnerability described by ...?'
• 'What time is it?'
• 'Calculate 15 * 8 + 32'
• 'What's the weather in Luxembourg?'

Chat session started. Type 'exit' to quit.

> What is the severity of a vulnerability described with: The Advanced Custom Fields: Extended plugin for WordPress is vulnerable to Privilege Escalation in all versions up to, and including, 0.9.2.1. This is due to the 'insert_user' function not restricting the roles with which a user can register. This makes it possible for unauthenticated attackers to supply the 'administrator' role during registration and gain administrator access to the site. Note: The vulnerability can only be exploited if 'role' is mapped to the custom field.    
╭──────────────────────────────────────────────────────────────────────────────────────── 🗨  tool_assistant@localhost/BFxpWUtCE0n3 ─────────────────────────────────────────────────────────────────────────────────────────╮
│ The severity of the described vulnerability is classified as Critical with a confidence of 58.26%. This indicates that the vulnerability poses a significant risk and should be addressed promptly to prevent             │
│ unauthorized access or privilege escalation.                                                                                                                                                                              │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
> exit

Chat session ended.

Agents are registered to the registry and presence notification system.

License

VulnAgent is free software released under the GNU General Public License version 3.

Copyright (c) 2025-2026 Computer Incident Response Center Luxembourg (CIRCL)
Copyright (c) 2025-2026 Cédric Bonhomme - https://github.com/cedricbonhomme