Module 16 · Chapter 10: Prompt-injection defence
Layered fences
In short
Wrap outside text as data through one door, with escaping and a rule in the system instructions. Then add layers outside the model: a reader with no tools, an allowlist, human confirmation, output checks and attack tests.
After this lesson you can
- Wrap untrusted text as data with a clear rule
- Build layered defences because one fence is never enough
Sign in to read this lesson
The title and summary are open to everyone. The full lesson is for signed-in learners.