Stage 2 · Lesson 6 of 7
Private data, untrusted content, a way out
Simon Willison named three capabilities that, together, let a steal data. He calls them the : access to private data, exposure to untrusted content, and a way to communicate externally, such as sending email, calling a web address or even rendering an image from a URL.
If one has all three, anyone who can put text in front of it can try to make it send private data out. Removing any one leg blocks this whole class of attack. That makes it one of the most practical design checks in AI security: list each agent's tools and mark which legs they give.
Common trapadding one harmless-looking tool, such as a web fetcher or image rendering, to an agent that already reads private data and .
A question to check yourself, an AI practice session and 1 cheat sheet entry come with the full roadmap. Unlock
Resources for this stage
Start with· follow one of these
- CoursePaidGrokking Prompt Engineering for Software Engineers
Design Gurus · Beginner · Subscription
How system prompts, context and structured output really work, the ground every LLM risk stands on
- CourseFreeHugging Face Learn
Hugging Face · Intermediate
Free courses on how language models, embeddings and agents work, at an engineer's level