The AI overlay your lockdown browser cannot see

What the overlay actually does

The candidate opens the locked exam, and on top of it sits a small window the browser cannot see. The overlay reads what is on screen, sends it to an AI assistant, and shows the answer. A keystroke summons it, a keystroke hides it. To the lockdown browser, nothing has changed: no tab was opened, no window was switched, the exam is still full screen and still has focus. The overlay is drawn by the operating system, above the browser, and the browser has no way of knowing it is there.

This is the new shape of the cheating problem. The tools most programs rely on were built to stop a student from opening a new tab, and the student of 2026 is not opening a new tab. Short videos now walk students through getting around lockdown browsers, and at least one popular lockdown browser is reported to run an AI overlay app inside it. The instructions circulate in candidate forums within days of any new exam launch, written for people with ordinary technical ability.

The lockdown browser is doing exactly what it was asked to do. It controls the exam window, and it does that well. The problem is that the exam window is no longer where the cheating happens.

Why the webcam only partly closes the gap

The usual response is to add video proctoring, live or recorded with automated flags. It helps with some of the problem and misses the rest. A webcam sees a face and a room. It does not see software. An overlay assistant, a remote session, or a virtual machine produces no visible change in the candidate's behaviour, and the video flags that remain useful are the ones that catch a second device or a second person, which is only a fraction of the problem.

Worse, the signals video can catch are exactly the signals a prepared candidate learns to avoid. Modern assistants are designed to sit in the candidate's line of sight, so the eyes do not move off screen. The candidate reads the answer where they would read the question. The recording shows a person looking at an exam, because that is what they are doing, while the answer arrives from elsewhere.

That does not make video useless. It makes it one layer, and a layer that covers the room, not the machine. A program that treats the recording as proof of a clean session is trusting the layer with the clearest limits.

What actually works now

Device-level monitoring is the first real answer: software that looks at what is running on the machine, not just what is in the browser. It can see the overlay process, the virtual machine, the remote session, and the AI tool, because it runs at the same level they do. A browser cannot see a phone; a device agent can see the process list. The two tools answer different questions, and the program needs both.

The second answer is exam design. Questions that ask for recall are the ones an overlay answers best. Questions that ask the candidate to apply, to reason from a scenario, to show working, are harder to answer from a floating window, and they are easier to flag afterwards, because the answer pattern looks different from the candidate's own short responses. A well-designed question is a detection tool in itself.

The third answer is the oldest one: move high-stakes assessments back to a controlled setting. An in-person exam or a proctored testing center removes the device problem entirely. It is not a step backwards. It is an admission that the remote device is now the candidate's device, and the candidate controls it.

Keep reading: What is a brain dump site, and how do certification programs fight them?, Your video feed dropped mid-exam: what the proctor does next, minute by minute, and What should a testing program do when AI writes the candidate's answers? cover the nearby cases in this series.