In many companies now, an employee asks the AI assistant a question. It searches the company’s documents and returns an answer in seconds. Staff love it, and it saves real time.
It’s also one of the easiest ways to leak your most sensitive information.
“Confidential” or “HR Only” does little to prevent an AI assistant from reading it and repeating its contents to someone who shouldn’t see it. The label acts as a sticky note, not a lock.
This catches leaders off guard because the old rules of data security taught us the label was enough. In a pre-AI world, security software watched the edges of the company network and flagged files based on tags and keywords. But once the AI is reading and summarizing your files, the risk is not at the edge of the network. It sits within the system, in determining which documents the AI is allowed to see.
The fix comes down to a handful of decisions about how the AI finds information and what you allow it to do. Here are the risks in plain terms, along with the questions that put you back in control.
Why labels disappear
An AI system breaks a document into small pieces so it can be searched quickly. In that process, information about who is allowed to see the document is usually stripped away. What remains is a pile of text fragments that have lost their link to the original permissions.
When an employee types “What are the severance terms for the Q3 layoffs,” the AI finds and returns the most relevant text. It doesn’t check the person’s title or clearance, just returns the closest match.
It’s a risky bet to assume the AI will catch it and politely refuse the request. A system designed to predict the next word can’t reliably enforce a hard rule about who sees what, because that’s not what it was built to do.
Real security means the AI never sees the off-limits document in the first place, using two methods. First, the AI should check permissions before searching, not after. If someone isn’t in HR, the severance file should simply not exist for their question.
Second, when the AI can take independent actions—such as sending an email, changing a record, or deleting data—it should be limited to a short list of pre-approved tasks, with every action logged.
6 decisions that turn a label into a lock
While you can’t configure any of this yourself, you can ask your team to verify that each is in place.
1. Permissions must travel with the file
When a document gets broken into pieces, those pieces tend to lose the original restrictions. If a file were limited to finance, every fragment of it should stay limited to finance. Require that permissions follow the document into the AI system, and that revoking someone’s access takes effect instantly, everywhere. A delayed update is an open window.
2. Block by default
Trying to scrub a secret from an answer after the AI has already written it is too late, because the information is already on the screen. Require the system to check permissions before retrieving anything, and return nothing when the person is not cleared. Silence is the safe default.
3. Judge access by the situation
Picture an employee with legitimate access logging in from a personal phone on airport Wi-Fi. Require rules that weigh things like the device and the location, not only the person’s role. The same employee can be trusted at their desk and blocked on an open network.
4. Test the system with planted secrets
The most common failure is assuming the controls work and never checking. Have your team plant fake, clearly marked secret documents, then run regular automated tests that attempt to retrieve them as an unauthorized user.
The only acceptable result is zero. If one planted secret ever surfaces, you have found a hole before a real document fell through it.
5. Keep a record of decisions
When something goes wrong, or a regulator comes asking, most logs show only the chat. Require logs that capture what was asked, which documents the AI used, and which it blocked and why. This also enables you to meet emerging standards such as the NIST AI Risk Management Framework and ISO 42001, the world’s first international standard for AI management systems.
6. Assign someone to high-stakes actions, and build a kill switch.
An AI that can act on its own, sending messages or changing records, can be confused or hijacked into firing off actions in rapid succession. Require human approval before it does anything costly or hard to reverse, and build a switch that instantly cuts the agent off the moment it starts behaving abnormally.
What the first 90 days should look like
You don’t need to follow the engineering to track the work. Expect three phases over the first three months.
- Month one: Your team takes inventory, finding every place where company documents were fed into an AI system and checks whether the permissions came along. If not, those documents get reloaded properly. The planted-secret tests get built now.
- Month two: The team switches on permission checks before retrieval, connecting the AI to your existing employee login system, so it always knows who is asking.
- Month three: Testing becomes automatic and continuous. From that point on, if a planted secret ever leaks during a test, the new version does not ship. Security stops being a one-time setup and becomes a standing check.
The bottom line
Treat security as something your AI must obey, not something you hope it respects. The goal is easy to state and hard to build. Your AI assistant should live within the same boundaries as the employees who use it.
When a file is marked confidential, it should stay confidential, no matter who asks. Get that right, and it becomes more than protection. It becomes a reason customers and partners trust you with their data.
Boris Dzhingarov is the CEO of ESBO Ltd.