You break your first LLM in two hours

You know detection. You know how to read a network, write a finding, run an engagement. Then a chatbot showed up in your ticket queue, or a client asked whether your pentest covers their assistant, or you shipped an agent yourself and later read the phrase indirect injection and went quiet. Here is the part nobody spells out. A language model cannot tell the difference between the instructions you gave it and the text it reads while doing its job. The system prompt, the user message, and the contents of an uploaded document all reach the model as one stream of tokens. That single fact is why a summarizer can be turned against you by a sentence buried in the file it summarizes, and why a support bot can be talked into printing the secret it was built to guard. You do not need machine learning to see this happen. You need a live target, a payload, and someone showing you where to type. That is the whole course. In two hours you attack five targets with your own hands, and you walk out able to test the assistant your own company shipped without taking down production.