The Lab
See every AI prompt your staff send
Frank De Pasquale ·
Someone in your business opened a browser tab this morning, pasted in a client contract or a list of customer records, and asked an AI tool to summarise it. Nothing about that felt like a security event to them. The tools are frictionless, they are quick, and they are useful, which is exactly why the usual habits about what you share with an outside system tend to quietly erode.
So it is worth asking what your security controls would actually do about that prompt. For most businesses the honest answer is nothing, and nobody would know it had happened. The material going into that text box is often the most sensitive the business owns. Draft agreements, pricing, source code, customer lists, health records. Close to 40% of what staff now put into AI tools includes sensitive data, and that share has climbed every year (Cyberhaven 2026).
Nobody gave staff a safe way to do this
So they found their own. There are more than a dozen of these tools in common use now, and people reach for whichever one is open. Close to half of generative AI use inside businesses runs through personal accounts rather than a managed corporate login, which puts it outside the company's identity and governance entirely (Netskope 2026). This is shadow AI, the same shape as the shadow IT problem before it and moving faster, with the number of incidents where sensitive data is sent to an AI tool roughly doubling in a year (Netskope 2026).
What almost no business has is a safe way to let staff use AI and still see what is happening. There is usually a policy that says be careful, and no way to tell whether it is being followed. That gap is the whole reason browser security matters here.
Where you can finally see the prompts
That AI text box sits in a browser tab, and a browser security layer (tekspace.com.au/solutions/browser-security) operates at exactly that point, where the data leaves. It records every prompt a staff member sends into a large language model, whether the login is personal or corporate, whether the tool is one of the well known names or one of the lesser known models a growing number of people are quietly using.
Those prompts land in one place your leadership can actually search: a running record of what has gone into every AI session across the business. That is the internal picture almost no organisation has today. From there the same layer can mask or block defined patterns before they leave the browser, so a card number or a patient identifier does not reach the prompt to begin with. The visibility comes first, and for most businesses the visibility is the piece that has been missing.
Visibility comes before control
You can't manage what you can't measure. So the first move we recommend here is never a purchase or a policy. It is a monitor-only look at what is already happening, usually across a small group of users, over a couple of weeks. Most leaders are surprised by what surfaces. It is usually the well intentioned staff member moving quickly, pasting something in to save an hour, with no sense that the material was sensitive or that anyone could see it.
Once you can see it, deciding what to do becomes a far easier conversation, with the evidence in front of you. We road test these platforms on what they surface in an environment like yours, and you can book a look at your own browser traffic or read the full analysis first.
Sources
1. Cyberhaven 2026 AI adoption/risk report — close to 40% of what staff put into AI tools includes sensitive data (39.7%; trend 10.7% > 27.4% > 34.8% > 39.7%) 2. Netskope Cloud & Threat Report 2026 — close to half of generative AI use runs through personal accounts, outside identity/governance (shadow AI ~47%; supersedes the 2025 edition's 72% figure, do not use 72% as current) 3. Netskope Cloud & Threat Report 2026 — incidents of sensitive data sent to AI tools roughly doubled in a year (~223 incidents/month/org avg, ~2x YoY)