APEX AI (Work in Progress)

APEX is an experimental, local-first agentic framework designed for cybersecurity automation. Built on LangChain4j 0.35.0, it utilizes the Jlama Vector Engine to run LLMs natively on the JVM, providing a secure environment for automated vulnerability research and network reconnaissance.

Important

This project is currently in Active Development. Features and internal APIs are subject to change as the agentic protocols are refined.

---

## 🛠 Project Vision

APEX is designed to bridge the gap between high-level reasoning and low-level system execution. By treating system operations (network scans, file I/O, web audits) as "skills," APEX can autonomously plan and execute multi-stage security tasks.

### Current Features (Implemented)

• Local Inference Engine: Native GGUF support via Jlama, optimized for Meta-Llama-3.1-8B-Instruct.
• Strategy-First Logic: APEX generates and critiques three distinct execution strategies before acting.
• Core Security Skills:
  • Network Recon: Automated nmap execution with dynamic path discovery.
  • Web Auditing: Programmable spidering and scanning via OWASP ZAP.
  • Local Intel: Integrated web research via SearXNG.
• Abstract Tasking: Decoupled architecture allowing tasks to be prioritized and managed as structured records.

---

## 🏗 System Architecture

The project is organized into several key packages:

• agentic.llm: Handles model provisioning and streaming configurations.
• agentic.tasks: Manages the TaskHierarchy, TriagePlanner, and strategic reasoning layers.
• agentic.tools: Contains the "hands" of the agent—classes like TerminalTools, ZapTool, and FileSystemTools that interact with the host OS and external security software.

---

## 🚦 Development Roadmap

• Phase 1: Foundation (Current)
  • Implement base LangChain4j service bridges.
  • Integrate local Jlama model loading.
  • Establish basic terminal and nmap toolsets.
• Phase 2: Enhanced Orchestration
  • Improve strategy critique logic (self-correction).
  • Expand ZAP integration to include report parsing and automated alert remediation.
• Phase 3: Multi-Modal Abstraction
  • Finalize "Abstract Data" types to seamlessly handle hardware sensor data and industrial protocols alongside standard IT logs.
• Phase 4: Persistence
  • Integrate a local vector store for long-term memory and CVE repository grounding.

---

## ⚙️ Installation (Dev Environment)

### Requirements

• Java 21 or 22 (Must support jdk.incubator.vector).
• OWASP ZAP (API enabled for local tool interaction).
• Maven for dependency management.

### Local Configuration

1. Configure Environment:

   # Enable Java Vector API for Jlama performance
   export JDK_JAVA_OPTIONS="--add-modules jdk.incubator.vector --enable-preview"

2. Clone & Build:

   git clone https://github.com/mbeds/agentic.git
   cd agentic
   mvn clean compile

3. Update LlmProvider.java: Set your local model cache and HuggingFace token in the getModel method to ensure persistence between reboots.

---

## ⚠️ Disclaimer

This is a work in progress intended for authorized security testing only. The automation capabilities within APEX can perform intrusive actions; always verify scan targets and ensure compliance with legal boundaries.

---

Current Version: 1.2.0-SNAPSHOT | Engine: LangChain4j 0.35.0