What a CISO actually asks in an AI security review

Most AI pilots in regulated firms do not fail on capability. From what we see, they fail in the security review, usually on a question the product team did not prepare for. These are the questions we hear most often, and the answers K-Lake gives.

How many copies of our documents will exist?

This is usually the first question, and the right one. Most AI tools work by copying content into a new index, often in a vendor's cloud. Every copy is something new to govern, retain and defend. K-Lake reads files where they already sit, on your shares, object storage and SharePoint, and keeps its index inside your infrastructure. There is no second copy of the corpus outside your perimeter.

Will the assistant respect the permissions we already have?

A CISO wants to know whether file permissions are enforced or re-created. Re-created permissions drift, and drift is a breach waiting to be found. K-Lake captures the permissions on each file at crawl time, maps them to your directory identities and enforces them in the database at query time. Access fails closed. If a person cannot open a file today, neither they nor an assistant acting for them will see it in an answer.

What does the model vendor receive?

If you connect a cloud assistant, it receives the specific passages retrieved for a question, already trimmed to the user's permissions, under your own agreement with that vendor. It never receives the original files. If that is still too much, K-Lake runs against a local model and nothing crosses the boundary at all.

Can we prove who asked what?

Every AI request and every file opened is recorded with the person, the assistant and the files returned, including requests that were refused. Logs and metrics go to your existing SIEM.

Can we switch it off?

The AI connection is disabled until you enable it. An administrator can end any assistant's session in one click, and can hide a sensitive source from AI while leaving it searchable for staff.

What about the software itself?

The core services and every connector are written in Rust, a memory-safe language, which removes whole classes of vulnerability by design. Our certification status is published on the trust centre and reviewed monthly.

If your security team has a questionnaire, send it to us before the proof of concept rather than after. It is the quickest way to find out whether we are a fit.