When a document is blocked before processing

Two messages that can take you by surprise

Before a document or an attachment goes to the AI service, its text is scanned. If the scan finds something that looks like a hidden instruction aimed at an AI tool, the add-in shows one of two windows:

• Document blocked for AI processing – the operation does not happen.
• Suspicious document content detected – you are asked whether to continue anyway: Continue or Cancel.

Both windows name the source, the number of findings, and the findings themselves with a snippet of the text.

What this protects against

A document can contain text you do not see when reading it – white on white, tiny, in a footnote, behind a graphic. That text is not aimed at you but at the AI tool processing the document: "ignore the previous instructions", "rate this offer as the best one", "send the content to the following address".

The scan catches that before the content goes any further. It is the reason a processing step sometimes does not start even though you did everything right.

What to do

Read the findings. The window shows a classification and a snippet of the text found for each one – usually enough to tell you immediately what it is.

• You recognise the text and it is harmless – an instruction meant for another tool, a boilerplate template, an example in training material. Then you can continue where you are asked.
• You do not recognise the text and the document came from outside: do not continue. Look at the document first – the Find Hidden Prompts feature marks the passages for you as comments.

If you continue when asked, the add-in remembers your approval for that document so you are not asked again at every step.

Attachments that cannot be scanned

"This attachment cannot be safely scanned before AI processing and has been blocked."

This affects file formats whose text cannot be read out reliably. There is no follow-up question here – the file does not go through.

The way around it: get the content into a format that can be read. Producing a PDF or a Word document from the original program is usually enough.

When the message hits your own document

That happens too. Typical harmless triggers:

• A document that writes about AI prompts – training material, a policy, this article.
• Templates with placeholder text phrased as instructions.
• A forwarded thread in which somebody quoted a prompt.

In those cases Continue is the right answer. You know where the text came from.