Overview
Exodus is an agentic penetration testing suite built to bridge the gap between large language models and real-world offensive security tooling. Rather than treating LLMs as passive assistants, Exodus gives them agency — the ability to plan, execute, and adapt multi-step attack chains using the full arsenal of tools available in Kali Linux.
Motivation
Traditional penetration testing requires deep expertise across networking, exploitation, and post-exploitation phases. While automated scanners exist, they lack the reasoning capabilities needed to chain findings together intelligently. Exodus was born from the idea that LLMs, when given the right tools and context, could replicate the decision-making process of a skilled pentester.
How It Works
The system operates through an agentic loop where the LLM:
- Reconnaissance — Gathers information about the target network using tools like Nmap, Whois, and DNS enumeration
- Analysis — Interprets scan results and identifies potential attack vectors
- Exploitation — Selects and executes appropriate tools from the Kali Linux toolkit
- Adaptation — Adjusts strategy based on results, pivoting when initial approaches fail
The LLM doesn't just run commands blindly — it maintains context across the entire engagement, understanding how each discovery relates to the broader attack surface.
Technical Details
- Agent Framework: Custom Python-based agent loop with tool-use capabilities
- Tool Integration: Direct interface to Kali Linux utilities (Nmap, Metasploit, Hydra, etc.)
- Context Management: Maintains a structured representation of discovered hosts, services, and vulnerabilities
- Safety Controls: Built-in scope limitations to prevent unintended lateral movement
Challenges
One of the biggest challenges was designing the tool interface layer. LLMs need structured, predictable output from security tools, but many Kali utilities have inconsistent output formats. Building robust parsers for each tool's output was essential for reliable agent behavior.
What I Learned
This project deepened my understanding of both offensive security methodology and the practical limitations of agentic AI systems. Prompt engineering for tool-use is fundamentally different from conversational AI — precision and structured reasoning are paramount.