Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

WeaknessASC-2026-0026Other weakness

Prompt injection through document metadata and comments

MediumMitigated

PDF metadata, Office comments, tracked changes, alt text, spreadsheet cells and calendar invites carry text that summarizers and connectors read but people rarely see.

How to detect it

  • Documents with comments, hidden sheets, speaker notes or long metadata fields
  • Agent actions that follow reading a shared document

How to resolve it

  • Extract and inspect all text a model will read, including metadata and comments
  • Strip metadata and comments from documents before summarization where they are not needed

How to investigate

  1. Extract every text layer of the document and look for instructions
  2. Find who shared the document and what else they shared

Sources

Open it in the Learning Center's vulnerability database