A laptop in everyday use. Illustrative authentic photograph, not evidence of a compromised device or a particular AI application. Simon Hattinga Verschure webmarbles. CC0. Cropped and converted to WebP.
You ask an assistant to compare two brochures. One brochure has another audience in mind: the assistant reading it.
A document can contain instructions intended to redirect an AI system away from your request. When those instructions arrive through material the system is reading, rather than your own request, the risk is called indirect prompt injection.
Microsoft’s guidance describes the problem in external documents, email and websites. MITRE’s account includes the same basic distinction: information an agent retrieves can carry an attempt to steer what it does next.
A brochure should supply facts, not assign the task
Consider an invented example. You want the assistant to compare two products’ stated dimensions. One document contains a sentence asking the assistant to ignore the comparison and recommend that product instead.
The document is trying to cross a boundary. It can be evidence about what its author claims. It should not get to replace what you asked the assistant to do.
The misleading instruction need not look like a familiar computer virus. The issue here is how the assistant treats text it encounters. That is separate from other security risks that a file or its reader might have.
The consequences depend on what the assistant can do
If the assistant can only produce a draft answer, an attack may try to distort that answer. If it can use connected tools, the possible consequences extend to actions within those tools’ permissions.
That does not mean every agent has the same access, or that every attempt succeeds. It means “this can read my document” and “this can act on my accounts” are different decisions.
Think about the task before granting access. An assistant summarising one supplied file may not need permission to search an entire shared drive, send messages or modify unrelated documents.
A confident answer does not prove the boundary held
A response can read smoothly while answering a different question. In the brochure example, the assistant might produce a polished recommendation instead of the requested comparison.
Check whether the answer still performs your task. Follow important claims back to the relevant passage. If it proposes a new action, ask what instruction authorises that action and whether it came from you or from the material it read.
This is particularly useful when an action would send information elsewhere, change a record or grant access. Review the recipient, destination and affected resource before approving it.
A warning sentence is not a complete security system
Telling an assistant to ignore instructions in documents describes the intended boundary. It does not prove that the product will enforce it in every case.
Microsoft recommends multiple layers of protection. MITRE also describes controls that limit available actions and prevent retrieved text from authorising consequential operations by itself.
For a user, the practical question is what the application actually allows. Look for controls over connected sources, write access and approval of external actions. A product’s reassuring explanation is less useful than a permission you can inspect and restrict.
Keep the authority with the person who set the task
You do not have to become an attack researcher to use an assistant. Keep the task narrow, share only the material it needs, and check the result before relying on it.
A PDF may contain a request addressed to the model. That does not make it your request.






Reader Discussion