Large Language Models are rapidly becoming part of the cybersecurity toolkit. Analysts use them for OSINT collection, threat intelligence reporting, and even offensive operations. But how well do we understand the tools we are adopting, and what happens when those same tools become the attack surface?
This hands-on, four-hour workshop takes participants through both sides of the LLM equation. Starting with prompting fundamentals and LLM foundations, participants will learn how to effectively use LLMs for security work. They will then apply those skills in practice: first using Claude AI integrated with Kali Linux via the Model Context Protocol (MCP) to conduct OSINT, generate threat intelligence reports, and hack a live target in the OffSec Proving Grounds Playground. Finally, the perspective flips entirely as participants learn to attack LLMs themselves through jailbreaking, prompt injection, improper output handling, and more.
This workshop bridges the gap between using AI as a force multiplier and understanding its vulnerabilities. Participants will leave with practical skills they can apply immediately, whether they work on a red team, blue team, or somewhere in between.
Workshop OutlinePrompting Fundamentals + LLM FoundationsThe workshop begins with practical prompting techniques for security work. Participants learn how to craft effective prompts that produce useful, actionable output rather than generic responses. This is immediately applicable regardless of which LLM they use in their daily work.
From there, we build the foundational understanding needed for the rest of the day: how LLMs generate output, why they hallucinate, what context windows mean for a pentest session, and the basics of responsible AI. This section is deliberately non-academic. The goal is to give participants just enough theory to understand why the techniques in later hours work and why critical evaluation of LLM output is essential.
OSINT & Threat Intelligence Reporting with LLMsParticipants shift from theory to practice, using Claude integrated with Kali Linux to conduct OSINT operations and produce structured threat intelligence reports. This section demonstrates the analyst-facing side of LLMs: how they can accelerate intelligence gathering, source analysis, and report writing.
Participants also learn to evaluate LLM output with the same rigor they would apply to any other intelligence source. What did the LLM find? What did it miss? What did it fabricate? This analytical discipline is what separates effective LLM-assisted analysts from those who blindly trust the output.
LLMs as a Hacking ToolNow participants use Claude and Kali Linux to hack a live target machine in the OffSec Proving Grounds Playground. Working through a full attack chain, they experience firsthand how an LLM can serve as a co-pilot during offensive operations: from initial enumeration and scanning through vulnerability identification to exploitation.
LLM Red TeamingThe perspective flips entirely. The LLM is no longer the tool; it is the target. Participants learn how to test and exploit vulnerabilities in LLM-powered applications, drawing directly from the OffSec LLM Red Teaming learning path. This section covers the techniques attackers use to manipulate, bypass, and abuse LLM systems.
Key Takeaways & Q&AThe final session brings everything together. We review key takeaways from all four hours, discuss where LLMs in cybersecurity are heading, and open the floor for questions and discussion.
Learning Objectives- Write effective prompts for security workflows and critically evaluate LLM-generated output
- Explain how LLMs generate output, why they hallucinate, and what this means for operational security work
- Conduct OSINT collection and produce structured threat intelligence reports using LLM-assisted workflows
- Use LLMs as a hacking co-pilot for enumeration, vulnerability discovery
- Identify and exploit LLM-specific vulnerabilities
