Pretty hard to fake an oral exam right now. But after a few generations of HUD improvements, even that might get tricky without sweeping the student for "bugs"
Have whisper listen to the question, feed it into chatgpt and read out the solution. The pipeline is already here
The software is there. Difficult to detect hardware is not (excluding low-bandwidth devices a la the Carlsen vs Niemann scandal)
The "controlled environment" may need to be a Faraday cage, among other things.
Even then, llama shows that local models can be fairly powerful. Tuning for, say, the GPUs used in handheld gaming devices, on a specific knowledge base, doesn't seem implausible.
That's true, and over the long term it's only going to get easier to hide these things. A shame, we might need to include metal detectors on the way in like they're in prison (and even that might not be enough).