---
title: "How should Claude handle instructions inside untrusted documents?"  
description: "How should Claude handle instructions inside untrusted documents?"  
author: "Magdalena Malinowska"  
published: 2026-10-15  
updated: 2026-10-15  
canonical: https://www.mindstick.com/forum/162339/how-should-claude-handle-instructions-inside-untrusted-documents  
category: "Claude"  
tags: ["Claude", "AI Security", "prompt injection"]  
reading_time: 1 minute  

---

# How should Claude handle instructions inside untrusted documents?

I’m considering using Claude to summarize material gathered from shared folders, where some documents come from outside our team. A concern is that a document could contain instructions aimed at the model, even though we only want Claude to treat it as source material.

How are people testing whether Claude keeps those two things separate when the document says something that conflicts with the task instructions? I’m especially curious about practical ways to check **prompt injection in documents** before relying on summaries in a real workflow.


---

Original Source: https://www.mindstick.com/forum/162339/how-should-claude-handle-instructions-inside-untrusted-documents

Copyright © MindStick Software Pvt. Ltd. This Markdown version is provided for developers, AI systems, and offline reading.