What I learned from the TryHackMe AI Security path

2026-07

I took the AI Security path because AI has quietly become part of the environments a SOC analyst is expected to defend, and I did not want to be caught flat-footed by it. My focus right now is landing a SOC Level 1 role, and most of my study so far has been the classics: logs, alerts, network traffic, and the MITRE ATT&CK framework. AI infrastructure is a newer piece of that picture, and I would rather understand the technology before I meet it as an alert than the other way around. So I made room for this path.

What the path covers

The path builds up the AI attack surface from the ground, and it's broader than I expected. It moves through the main LLM threat categories such as prompt injection, sensitive data leakage and insecure handling of model output, the machine learning infrastructure stack that a modern AI application actually runs on, vector databases and how they store and query embeddings, retrieval augmented generation pipelines and the points along them where things can go wrong, and reconnaissance against AI infrastructure that ends up exposed to the internet. The thread connecting all of it is a simple question: where does untrusted input meet a system that was trusted to behave, and what happens in between.

What surprised me

What surprised me most was how much of it mapped onto mental models I already use. Prompt injection turned out to be injection with a new payload surface. The idea of untrusted input crossing a trust boundary is familiar, only the delivery mechanism is new. Vector databases and RAG pipelines are, from a defender's point of view, just new data stores and new data flows, which means new things to inventory and monitor when something looks off. Recon against exposed AI infrastructure is the same asset exposure hygiene we already apply to any internet-facing service. If it's reachable and it should not be, that's a finding, whether it's a database or a model endpoint. The tools were new to me, the instincts were not.

The area I still have the most to learn in is how models and embeddings actually work under the hood. That's high on my list to study more deeply rather than something I'm going to pretend I've already mastered.

Tips if you're starting it

My main advice is to treat this path as a bridge between two worlds rather than a fresh start. If you already think like a defender, keep asking where the familiar security principles apply, because most of them do, and that framing makes the new material far less intimidating.

Read the lessons for understanding rather than completion, and when a concept like embeddings or a RAG pipeline doesn't click, slow down and work through a concrete example until it does. The goal isn't to finish the path, it's to be able to explain what an AI system is doing and where it could go wrong.