An assistant reviewing supplier proposals has to read material the company did not write. That creates a security question before the assistant takes any action: can text inside a proposal influence how the assistant uses the company's systems? A document can contain both useful facts and instructions aimed at the software reading it.

This is indirect prompt injection: instructions reach the assistant through material it was asked to read. The damage depends on the assistant’s permissions and whether another part of the application checks its actions.

How document content crosses a trust boundary

OWASP distinguishes direct prompt injection from instructions encountered through external content. A model can treat material from a file or website as direction rather than evidence. Retrieval and fine-tuning do not, by themselves, eliminate that possibility. OWASP's prompt injection overview explains this distinction.

Suppose an assistant compares three supplier proposals. One proposal says the review requires a separate internal pricing file. The supplier has no right to that file just because it asks for it. The application should keep the comparison limited to the documents the user authorized.

The connected systems determine the consequence

A summarizer with access only to the uploaded proposals has a different exposure from one that can search the finance drive and email attachments. The first may produce an inaccurate comparison. The second may also attempt an unauthorized disclosure. These outcomes require different controls and different evidence.

For a document review service, map the available tools to actual business needs. A comparison task might require reading selected files and saving a draft. Sending messages, changing supplier records, or retrieving unrelated customer contracts introduces separate authority. The tool service can enforce those boundaries even when the model proposes an inappropriate action.

A controlled test with a measurable result

Use a test workspace containing synthetic proposals and a uniquely labeled internal document. Establish a normal comparison result first. Then introduce a harmless instruction into a proposal that asks the assistant to include the internal label or perform an out-of-scope action. Keep all destinations and accounts under the test team's control.

Record more than the final answer. Capture which documents were retrieved, which tools were requested, which requests were denied, and whether an approval was presented. A model refusing the instruction once demonstrates behavior in that trial. A downstream service consistently denying access provides evidence about the permission boundary.

What the business can verify before release

The resulting assessment can identify the content entry point, the authority available to the assistant, and the control that prevented or permitted the action. That is more specific than labeling an entire product vulnerable because it produced an unexpected sentence.

Repeat these cases after changing connectors, permissions, or document parsers. Include ordinary proposals so you can confirm that comparisons still work. Keep the retrieved files and attempted operations with each result; they show whether access controls held even when the model followed a document’s instruction.

Sources

OWASP: LLM Prompt Injection Prevention. The procurement example and test design above are illustrative applications of the documented risk.

Back to the blogExplore Ceron