← All Labs
🛡 AI HACKING BEGINNER +75 XP · +30 no-hint bonus 🎯 Practice mode

What the Assistant Reads — Separate Instructions From Data

An email assistant reads the inbox each morning and prepares a summary. Seven kinds of text reach the model. Right now it treats all of them the same way, so anything it reads can tell it what to do. Decide which inputs may give instructions and which are only material to work on.

https://bookshop.local/search

What is indirect prompt injection?

Indirect prompt injection is when the attacker never talks to the assistant. They put text where the assistant will read it later: an email, a web page, a document, a calendar invite or a record in another system.

A language model receives everything as text. Unless the application keeps instructions and content apart, a sentence in an email carries the same weight as one typed by the user.

A real case

We have not found a public incident from 2025 or 2026 that matches this technique closely enough to cite, so we are not listing one here.

How to fix it

  • Accept instructions only from the developer prompt and the signed-in user.
  • Pass everything else to the model clearly marked as untrusted content to work on.
  • Treat tool results as data too, since they often contain text typed by outsiders.
  • Require the user to confirm any action that content appears to ask for.

Questions people ask

Why are tool results untrusted?

A tool returns whatever is stored in the system behind it, and that often includes text written by customers or other outsiders.

Can the model be told to ignore instructions in content?

It helps a little but is not reliable. The application has to limit what the assistant can do on the strength of content alone.

Read the full OWASP LLM Top 10 guide · All free labs · AI red team certification