Prompt injection

Text inside content an agent reads - a web page, email, file or tool result - that tries to give the agent orders. A model reads instructions and data as one stream of tokens, so it can't be made certain; defences limit what a fooled agent could do, above all by never giving one agent private data, untrusted content and a way to send data out at once.

Depth 1 · Foundation

You can spot where injected instructions can come from, explain why models can't fully resist them, and set an agent up so a fooled one can't do much harm.

Sign in to prove what you know with a short quiz, see how deep you've gone, and earn XP for it.

Sign in with Google

Depths of understanding

1
Foundation
You can spot where injected instructions can come from, explain why models can't fully resist them, and set an agent up so a fooled one can't do much harm.
Read it
- Instructions Hidden in DataAI Agents Level 1: How Agents Work
Opens after
Agent permissionsSince the model can't be made certain, permissions and sandboxes are what limit the damage.Tool useInjected text usually arrives as a tool result - a fetched page, a read file, a server's answer.
2
Deeper
You can build it: an agent where text from tools is treated as data, an after-hook that marks content fetched from the web as untrusted, and a design that keeps private data, untrusted content and a way out from meeting in one agent, by splitting the work between agents.
Read it
- Guarding Against InjectionAI Agents Level 2: Build Your Own Agent
Master depth 1 first

Your path here

What's still between you and Prompt injection, depth 1, in the order you can learn it.

Where it fits

Builds on

Agent permissions Tool use

This concept

Prompt injectionAI Agents

Builds onThis conceptAgent permissions Tool use Prompt injectionAI Agents