We spent months trying to cheat our own interview system

Before we let a single client near the interview platform, we spent several months attacking it.

Not in a formal, write-up-a-threat-model way. More like: sit down, open a session, and try to get through it dishonestly. I did most of them myself. I have now taken more interviews in my own product than anyone else ever will, and a good number of those were me trying to beat it.

What we actually tried

The list got long. A phone propped below the camera, angled so the eyes barely move. A second monitor with an LLM open on it. Bluetooth earbuds with someone else reading answers into my ear. Lip-syncing to a person speaking off-screen. Switching browser tabs. Pasting in generated text. Asking someone to sit just outside the frame and prompt me. Virtual cameras. Audio played quietly in the background. And the low-tech one that people underestimate — simply stalling, saying nothing for a long stretch while something else worked on the answer.

Some of these died instantly. Others survived much longer than I was comfortable with.

The one that got furthest

The second screen with a language model on it. Not because it is clever, but because it looks like nothing. A disciplined candidate reading off a second display does not look furtive. They look like someone thinking.

But the thing that actually unsettled me was not the setup. It was the realisation that a fairly ordinary candidate, with a small amount of outside help, could look considerably better than they were. Not brilliant. Just good enough to get through a screen and into a room with a hiring manager who would then have to work out, in forty-five minutes, what a whole process had failed to. That is the real cost of a cheatable interview, and it lands on someone else's calendar.

What surprised me

I assumed we would spend most of the effort on the obvious things. Eyes drifting off-screen, someone else's voice, a second face in the frame. We did, and those are largely solved problems.

What I did not expect was how quickly a borrowed answer stops holding up once you push on it. The final answer can be perfectly correct. It is everything around the answer that gives way — the moment you ask why they ruled out the other approach, or what happens to their design when one assumption changes, or you take a detail they mentioned in passing and build the next question out of it. Assistance is good at producing answers. It is much worse at sustaining a position across ten minutes of pressure it did not see coming.

That changed how I think about the problem. We had been treating this as a detection question. It is at least as much a design question.

What our friends found

At some point we stopped guessing and asked people to break it for us. Friends and family, no brief, no hints about how it worked. Just: get through this without doing the work.

They found things we had not considered. Most of it clustered around the physical setup rather than the software — where exactly a camera can be placed, what you can do with audio hardware, how much of a second person you can keep just outside the visible frame. Several of those went straight into the test set and stayed there.

I would recommend this to anyone building anything in this space. You will not think of the ideas that someone with no investment in your assumptions will think of in ten minutes.

What I am still not confident about

No system catches everything, and I would not trust anyone who says theirs does.

The hardest case is not the elaborate setup. It is a well-prepared candidate using a small amount of assistance, rarely, at exactly the right moment. Sparing and disciplined is much harder than constant and greedy. I do not think that case is fully solved, by us or by anyone.

Which is why proctoring on its own is not the answer, and why I am wary of platforms that lead with it. Watching for cheating is worth doing. But an interview that is hard to fake in the first place — because it follows you, and asks about the thing you just said, and does not accept an answer as finished — does more work than any amount of monitoring bolted onto an interview that was easy to fake to begin with.

We ran roughly a hundred controlled sessions over several months. Every time something got through, we treated it as unfinished rather than unlucky. That is still how we work on it.