Your firewall stops traffic. It does not stop words. In Part 2 of our Domain 2 Deep Dive, we cover the most dangerous (and fun) part of AI Security: Input Attacks.

We explain how attackers use "Prompt Injection" to turn helpful chatbots into "Confused Deputies" that attack their own users. We break down the difference between Jailbreaking (Roleplaying/DAN) and the mathematical magic of Universal Adversarial Triggers (UATs).

But the scariest attack isn't when you talk to the AI—it's when the AI reads a file you didn't check. We demonstrate Indirect Prompt Injection and how a simple resume PDF can hack your recruiting bot.

🎓 In this video, you will learn:

Context Mixing: Why LLMs cannot distinguish between "Data" and "Instructions."

Jailbreaking: "DAN" (Roleplaying) vs. "Logical Bypasses" (Translation exploits).