AI

Every security club now has an argument about this, so here is the club’s position up front: these tools change what one person can do in an afternoon, and they are a new attack surface. Both statements are true at the same time. Neither one cancels the other out.

This section covers what a large language model (LLM) is, what an agent system is made of, what runs the loop, and how knowledge gets in front of one. It covers what happens when the same technology is pointed at you. Then the practical half: which model to pick, how to wire one into a script, and what never goes in a prompt.

Why This Is Here

Three reasons.

It changes the leverage of one person. A student with a lab VM, a shell, and a model that can read a thousand log lines is doing work that used to need a team. That is real. It is also the exact reason your fundamentals matter more than they did before — you cannot check work you do not understand.

It is a target. An agent that reads a web page, a ticket, or a log file is reading text an attacker can write. The model has no built-in separation between “text to process” and “instructions to follow.” That single fact generates a whole class of attacks, and a cyber club is exactly the place to learn it.

It is already on both sides. Defenders use it for triage and summarization. Attackers use it for recon and for finding bugs. Neither side gets to opt out, so learn how it works.

You Are Still Accountable

IBM put it on a training slide in 1979: a computer can never be held accountable, therefore a computer must never make a management decision. Nothing since has made it less true.

A model will draft the plan, write the script, and read the logs for you. It will not answer for any of it. When the pipeline wipes the wrong target, when the agent follows an instruction buried in a page it was told to read, when a summary quietly drops the one line that mattered — that lands on you. The tool is not in the room when it goes wrong.

So the rule underneath this whole section: stay in control of your systems and understand what is happening inside them. Not trust the model, not trust the output — understand it well enough to catch it when it is wrong. It is the same reason your fundamentals matter, from the other direction: a decision you cannot check is a decision you have already handed away.

• Fundamentals for the AI Era What to actually learn now that models write code: the framings worth knowing, the case for small local models, and why the rest of the roadmap matters more.
• Agentic Graphs Nodes, edges, state and cycles: the vocabulary for describing systems built from many model calls, the named patterns, and the failure modes.
• Harnesses & the Agent Loop What actually turns a language model into an agent: the loop, the harness around it, and the tools people really use.
• Knowledge Bases as Agent Context Knowledge graphs against vector retrieval in plain terms, and what document structure actually helps an agent — with this wiki as the worked example.
• AI on Both Sides Defensive and offensive application of AI, and the part a cyber club uniquely needs: how agent systems themselves get attacked and defended.
• Choosing an AI Model Selection criteria for picking an AI model for security work — not a product comparison.
• AI Workflows & Agents The agent loop, common agent patterns, and vendor-neutral workflows for using AI in security work.
• Automating Security Work How to wire a model into a repeatable pipeline for triage, summarizing, and first-pass review — and what never goes in the prompt.