Hugging Face's logo

automajicly
/
Local_Model

Text Generation
English
mcp
security
autonomous-agent
ai-agent
cybersecurity
local-llm
Model card Files
xet
 Community
Local_Model / README.md
automajicly's picture
automajicly
Update README.md
469397a verified
preview code
|
raw
history blame contribute delete
3.3 kB
metadata 
license: apache-2.0
language: en
pipeline_tag: text-generation
tags:
  - mcp
  - security
  - autonomous-agent
  - ai-agent
  - cybersecurity
  - local-llm

Autonomous Security Agent

A self-contained security agent built with Qwen 2.5-7B running locally via LM Studio on Kali Linux. The agent can autonomously execute security tools, analyze results, and take action through an MCP (Model Context Protocol) server. Built for local offensive security automation with real-time tool execution and autonomous decision loops.

System Overview

This agent operates as a fully local autonomous security system that reasons, plans, and executes security tooling through a structured MCP execution layer.

It continuously cycles between:

  • interpreting objectives
  • selecting tools
  • executing system-level security commands
  • analyzing results locally via an embedded LLM

Features

A modular autonomous security system built around a local reasoning engine with executable tool orchestration via MCP.

  • Local LLM Backend — Qwen 2.5-7B served via LM Studio at 192.168.0.39:1234
  • Autonomous Tool Execution — Runs security tools (nmap, masscan) through MCP
  • Agent Loop — Continuous reasoning and decision-making
  • MCP Server — Tool chain execution with run_masscan, run_nmap, write_file, read_file

Components

System Architecture

The system is composed of four interacting layers:

1. Reasoning Layer

  • agent_loop.py — Core autonomous decision engine
  • Handles planning, evaluation, and iterative reasoning cycles

2. Execution Layer

  • mcp_server.py — Model Context Protocol execution backend
  • Routes structured tool calls into system commands

3. Tooling Layer

  • tools_manifest.json — Defines available actions and capabilities
  • run_nmap, run_masscan, read_file, write_file

4. Interface Layer

  • request.json — Input schema for agent initiation
  • Defines structured prompts and task definitions

Security Setup

Firewall Configuration

  • Default policy enforces outbound allow / inbound deny isolation
  • Suricata deployed for behavioral intrusion detection and alerting
  • System operates in execution-restricted network mode during agent runtime

Network Security Model

  • TOR routing available for privacy-preserving traffic paths
  • All LLM inference remains fully local (no external API dependency)
  • MCP server restricted to localhost-bound execution only

Installation & Setup

  1. Prepare a Kali Linux environment with network security tooling (Suricata enabled)
  2. Load Qwen 2.5-7B into LM Studio and expose local endpoint
  3. Configure system firewall rules for isolated execution mode
  4. Clone this repository
  5. Install Python dependencies
  6. Run the agent: python agent_loop.py

Documentation

See the docs/ folder for:

  • Detailed setup instructions
  • Firewall rule examples
  • Suricata configuration
  • MCP server setup

System Behavior Loop

At runtime, the agent follows a continuous operational cycle:

  1. Receive objective via structured input
  2. Interpret intent using local LLM reasoning layer
  3. Select appropriate tool(s) from MCP registry
  4. Execute actions through MCP server
  5. Collect and normalize results
  6. Re-evaluate state and iterate until completion

License

MIT