All the notes here are taken from an interview with Jason Haddix
Prompt Injection
The strategic manipulation of input prompts with an intent to bypass the LLM’s ethical, legal or any other forms of constraints imposed by the developers
Breaking down the primitives of prompt injection into 4 categories
Intents
Things that you propose to do to your AI system to hack it.
.
Techniques
Things that help you achieve your desired Intent
Evasion
Methods to hide and obfuscate your attack. Things like Obscure languages, truncated “wrds,” pig-latin, morse code, emoji metadata, etc
