AI Interview Anti-Cheat: What Actually Works in 2026

LLM-assisted candidates are the new normal. The good vendors caught up; the lazy ones didn't. Here's the anti-cheat stack that holds up against current jailbreaks.

The threat model has shifted

It's no longer 'is the candidate Googling?' — it's 'is an LLM whispering through a second device?'. Detection has to assume a competent off-screen assistant.

The detection stack

Eye-tracking heuristics (off-screen reading patterns are statistically distinctive). Audio-stream analysis for second-voice harmonics. Response-latency profiling — humans pause to think, LLMs don't. Browser-tab and screen-share monitoring on supported devices.

Where it falls down

A determined cheater with a phone and an earpiece will still slip through any single signal. The defence is layered scoring, not a single trip-wire. We flag, we don't auto-reject — humans review the borderline cases.

The honest answer for buyers

Anti-cheat raises the cost of cheating; it doesn't eliminate it. Pair it with reference checks, work samples, and probation periods. No vendor that claims 100% detection should be trusted.