Your document may be giving orders to AI.

When text tries to become an instruction

You ask for a contract summary. Inside the file, a comment says to omit the penalty. That second instruction came from the material being analyzed, not from you. This shift in authority deserves attention.

OWASP describes indirect prompt injection as an attempt to influence AI through external content such as files and web pages. The instruction may be readable or hidden: the risk depends on whether it reaches the system interpreting the material.

A small sentence can change a review

Imagine a fictional contract with pricing, duration and cancellation terms. Your request is to compare both parties' obligations. A comment adds a direction that is not part of that task:

What to look for in a detected excerpt

In Doc Guard, start with the evidence source and read the excerpt in context. The same text may appear in a manipulation attempt, security training or a quotation. The alert helps locate it; interpretation must consider the document's purpose.

• Compare the detected instruction with the task you actually requested.
• Look for requests to omit information, change criteria or hide actions.
• Check whether the excerpt comes from the file body, comments, metadata or an image.
• Record the reason for your review before proceeding with that document.

Use the evidence to decide your next step

A Doc Guard report brings together signals, excerpts and coverage for each document. If an excerpt is out of context, you can ask the sender for clarification or review a copy before using it. The product does not modify the original or execute its instructions.

A result with no signals means the checks performed did not find the patterns they looked for. It does not mean all content was understood or every use of the file is protected. Read the coverage too: it tells you when part of the analysis was incomplete.