Series1 part · In progress
Grokking AI Security
Working through AI Security Bootcamp one day at a time, with the exercises and what they show about attacking and defending LLM systems.
A series where I work through AI Security Bootcamp one day at a time. Each part covers the exercises from one day and what they show about how LLM systems can be attacked and defended.
Parts
- Part 1
Everything is One Token Stream
An LLM sees one stream of tokens. All the different components—system prompt, user messages, tool results, model responses—serve different purposes, have different trustworthiness, and carry different authority, yet they all appear in one token stream.