SecureLLM.ai

Before you put AI on your company's documents

Could an AI assistant leak your company's information?

Your legal team has said information can't leak between departments, or be exposed by accident. These short guides explain, in plain English, how an AI assistant could do both, what to ask before you approve one, and how a well-built system prevents it.

Three departments, HR, Finance and Sales, kept apart by walls, with one dashed line crossing from HR into Sales: the leak these guides are about. HR Finance Sales ?

The short answer

Can it really leak our information?

Yes, if it's built carelessly. It can put text from files a person was never allowed to open into the answers it gives them. Built properly, it follows the same access rules your files already have, and checks them before it reads anything.

The everyday picture: your office has a keycard system. Your badge opens some doors and not others, and your files work the same way. A well-built AI assistant carries each file's keycard rules with it and checks your badge before it fetches anything. A careless one is a helpful colleague holding a master key.

Where the picture stops working: an answer can blend passages from many files, so the check has to happen for every passage, not once at the door.

Engineers call these assistants RAG, short for retrieval-augmented generation: the AI looks things up in your documents ("retrieval") before it writes the answer ("generation").

Why this matters

Security is where AI projects stall

76%

of IT and security leaders named data security and privacy their top concern about generative AI.

87%

had delayed a rollout over data security and how information is managed. The delays were about readiness, not budget.

Source: AvePoint, The State of AI 2026, a survey of 750 IT and security leaders. The survey's sponsor sells data-security software.

Where leaks happen

From your files to an answer: where information can slip out

An assistant of this kind typically takes your documents through a few steps before it answers. Each step is a place where the keycard rules can be lost. Read them all, or walk through them one at a time.

  1. Your documents

    What you already have, in SharePoint, Google Drive or your document system, each with its own rules about who can open it.

    Can go wrong: many files are shared more widely than anyone realizes. Nobody noticed, because nobody could find them. An assistant can find them in seconds.

    Ask: have we checked who can open what, before we connect it?

  2. Bring in and prepare

    The assistant takes in your files and cuts them into passages it can look up.

    Can go wrong: if each file's keycard rules aren't copied in with it, the assistant can't respect them later.

    Ask: does it copy the access rules for every document, and how quickly does it notice when they change?

  3. Build the catalogs

    It builds catalogs, like the index at the back of a book, so any passage is quick to find.

    Can go wrong: if every department's documents go into one catalog with no walls, one search can reach all of them.

    Ask: how are departments, or clients, kept apart inside the system?

  4. Find the right passages

    When someone asks a question, it searches the catalogs for the passages that best match.

    Can go wrong: if it finds the best matches first and checks permissions afterwards, restricted text can slip through.

    Ask: does it check who is asking before it reads anything?

  5. Answer, with sources

    The AI writes an answer from those passages and shows where each part came from.

    Can go wrong: a summary, a quote or a source link can give away what's in a file the person couldn't open. A document can also carry hidden instructions that the AI follows.

    Ask: what stops an answer revealing something the person couldn't open?

  6. Check and keep current

    The system is checked for quality and kept up to date as documents change.

    Can go wrong: without a record, nobody can say who saw what. A deleted file can live on in saved copies.

    Ask: can we see who asked what, and what they were shown? When we delete something, is it gone everywhere?

  7. Put it where people work

    In chat, on the intranet, in email, or connected to other systems so it can act, not just answer.

    Can go wrong: once it can send emails or update records, it matters whose permissions it acts with.

    Ask: whose access does it use when it does something for someone?

Who we are

Written by people who have built one

We built Arro, a platform that answers questions from a company's own documents. Keeping each person's answers inside what they're allowed to see is part of that work, and these guides come from it.

  • Plain English first. Technical terms only when you'll hear them in a meeting, and always explained.
  • Every fact has a source you can check.
  • We never rank or score our own products. Where we describe how we built something in Arro, it's labeled.

When you're ready

Need an AI project through security review? Talk to the people who wrote this.

Moves and Measures helps companies plan and build AI assistants on their own documents, with the access rules done properly from the start.

Email Moves and Measures

contact@movesandmeasures.com