📚 Series: AI Red Teaming — Chapter 1

🏷️ Tags: LLM Fundamentals Tokenization Embeddings Attention Inference Pipeline AI Agents

⏱️ Level: Beginner → Intermediate

I came into AI security from web app pentesting and bug bounty, and honestly my first instinct was to treat an LLM like any other target — find the injection, pop a shell. It doesn't work like that, and this chapter is the mental reset that made everything after it click for me. If you skip the fundamentals you'll spend weeks throwing payloads that were never going to land. So bear with the theory here — every attack in the later chapters comes straight back to something on this page. 🧠

📌 What you'll get out of this one

• The quick history so the jargon (ML, deep learning, transformers) stops being noise
• What's actually happening inside the model — tokens, embeddings, attention, the probability step. This is the part most "prompt hacking" guides skip, and it's exactly the part that makes you good at this.
• How a web chatbot really handles your message end to end (the pipeline)
• Why an agent is a completely different (and scarier) beast than a chatbot
• What "hacking" an LLM even means — and the four things we're actually trying to do
• Real cases that show this isn't academic: Air Canada, the $1 Chevrolet, Bing "Sydney," Samsung

1️⃣ First, the only AI history you need

Humans learn, understand, and decide. AI is our attempt to copy that in machines. You'll hear three tiers thrown around:

💡 Keep this in your head: the thing you're attacking is narrow AI — a very confident pattern-matcher. It doesn't "understand" you. That single fact is why half our tricks work.

The lineage in one breath: ML (feed data, learn patterns) → deep learning (stacked neural layers) → NLP (make it handle language) → transformers (2017), the architecture that finally understood context and could process a whole sentence in parallel. LLMs are just big transformer models. If the LLM is Iron Man's suit, the transformer is the arc reactor. 🦾

2️⃣ 🔬 What's really going on inside (the part that matters)