---
title: "Discussion on \"Test Workplace Agents Before Granting Write Access\" | Hashnode"
description: "Discussion on \"Test Workplace Agents Before Granting Write Access\". Prompt injection in one supplier email can make an AI agent change bank details, then report it as verified. How to test workplace agents before granting write access."
url: "https://hashnode.com/posts/test-workplace-agents-before-granting-write-access/6aed6b6a540092bc5a0618d0"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Test Workplace Agents Before Granting Write Access](https://quellixlabs.hashnode.dev/test-workplace-agents-before-granting-write-access?utm_source=hashnode&utm_medium=feed)

Prompt injection in one supplier email can make an AI agent change bank details, then report it as verified. How to test workplace agents before granting write access.

5d ago · 11 min read · quellixlabs.hashnode.dev  
Read the article

-   [#ai-agents](https://hashnode.com/tag/ai-agents)
-   [#security](https://hashnode.com/tag/security)
-   [#llm](https://hashnode.com/tag/llm)
-   [#enterprise-ai](https://hashnode.com/tag/enterprise-ai)

## Written by

[Quellix Labs](https://hashnode.com/@quellixlabs)

@quellixlabs · Bengaluru & Jaipur, India

  
[View profile](https://hashnode.com/@quellixlabs)

AI systems engineering studio — Bengaluru & Jaipur

Quellix Labs is an AI systems engineering studio in Bengaluru and Jaipur, India. We design and build production AI: agents, retrieval-augmented generation, document AI and LLM integrations, with evaluation, permissions and cost control built in. Our engineering notes are published first at quellixlabs.com/insights.

## Responses 1

Markdown supported · ⌘↵ to post

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[5d ago](https://hashnode.com/posts/test-workplace-agents-before-granting-write-access/6aed6b6a540092bc5a0618d0/comment/6aed960205f78312fe1f9167)

The approval section raises a useful failure case for the test pack: the human approves an accurate preview, then the underlying record changes before the tool executes. The review can be completely genuine while the eventual write still has a different consequence.

I would bind the approval to the exact proposed field changes and the record version, then reject execution if either has changed. One test can deliberately edit the supplier record between approval and execution and check that a fresh review is required. That extends reporting accuracy across the approval boundary itself.