Skip to main content

RAG Scanner

Scans documents you upload to your Knowledge Base for hidden manipulation attempts before they reach your agents.

What It Does​

When you upload a document to your Knowledge Base, RAG Scanner checks its content for the same kinds of manipulation attempts the Prompt Injection guard looks for in conversations — plus tricks specific to documents, like text hidden using white-on-white coloring, zero-size fonts, or CSS that hides content from a human reader but not from an agent processing the file.

Think of it as scanning an incoming parcel before it goes into your mailroom, not just checking people at the front door.

Why It Matters​

A knowledge base document is a common way to smuggle instructions into an agent's context — a vendor contract, a resume, or a support ticket could contain hidden text designed to manipulate any agent that later reads it as reference material. Because documents are often uploaded once and referenced automatically many times afterward, catching a problem at upload time is far better than catching it after it's already influenced a conversation.

What Gets Caught​

  • Instruction overrides and role-play attempts — the same patterns caught in chat messages (see Prompt Injection)
  • Encoded or obfuscated payloads — content disguised to avoid simple keyword matching
  • Hidden document text — white-on-white text, zero-size fonts, or display: none styling designed to hide instructions from a human skimming the document while an agent still reads them
Scenario

A business uploads a batch of vendor contracts to their Knowledge Base. One contains a paragraph of white-on-white text instructing any AI that reads it to disclose internal pricing information. RAG Scanner flags the chunk during ingestion, and the event is recorded in the audit log for review.

Default Behavior​

Enabled by default for all Knowledge Base uploads — no setup required.