TL;DR — Generative AI has quietly turned the modern remote technical interview into a take-home test where the candidate isn't the one taking it. If you haven't changed how you interview since 2022, you are almost certainly hiring people who can't do the job. This is a practical 2026 playbook for engineering leaders who need real signal, fast.
The interview you thought you were running is gone
You schedule a 60-minute remote coding interview. The candidate shares their screen. They talk through the problem calmly, write clean code, and even catch an edge case you only hint at. You end the call, type "strong hire" into the scorecard, and move them to the final round.
Two weeks later they join the team. Three months later you let them go.
If this has happened to you in the last twelve months, you are not alone — and you are not unlucky. You are hiring in an environment that fundamentally changed between 2022 and 2026, and the interview process at most scale-ups has not caught up. Large language models are now so capable, so fast, and so well-integrated into the operating system that a competent coder can appear brilliant, and an incompetent one can appear competent — with nothing more than a second laptop and a prompt.
This post is for engineering leaders who feel that something is off with their pipeline but cannot yet name it. We'll walk through what AI cheating actually looks like in 2026, why the usual fixes don't work anymore, and what a defensible interview loop looks like going forward.
What "AI cheating" actually looks like in 2026
A year ago, "AI cheating" meant a candidate alt-tabbing to ChatGPT and pasting code back into a CoderPad tab. That was the amateur version, and it was easy to spot because the candidate had to pause, their eyes would move off-screen, and the code style would change mid-problem.
The 2026 version looks nothing like that. Here are the patterns engineering leaders are reporting in the field.
Second-device stealth. The candidate keeps a second laptop or phone just out of camera frame. A real-time audio-to-LLM pipeline listens to your question, generates a spoken answer, and feeds it back through an earpiece. The candidate is, in effect, a human bridge between you and the model.
Screen-overlay assistants. Browser extensions and native apps render model output as a translucent overlay on the candidate's own screen — invisible in any screen share. The candidate is reading from an autocue that you cannot see, even though you believe you are watching their full screen.
Voice and face replication. For the most contested senior roles, we are now seeing candidates who pass the phone screen flawlessly and then send a completely different person to the onsite — because the "candidate" on video was a real-time deepfake puppeteered by someone else. This is still the tail of the distribution, but it is a growing tail.
Proxy interviewees. Entire cottage industries have emerged around "interview-as-a-service," where a senior engineer in one country takes interviews on behalf of a mid-level engineer in another, splitting the first-year comp. The resume, the LinkedIn, the face on the call — all real. Just not the same person who will show up to stand-ups.
The common thread: in every case, the person you think is being evaluated is not the person you are evaluating. And the default modern interview stack — LeetCode-style problems on a shared editor, behavioral questions over Zoom, a system design round on a whiteboard app — assumes a world where that is impossible. That world no longer exists.
Why the obvious fixes don't work
When CTOs first become aware of the problem, the reaction is usually some combination of the following. None of them work on their own in 2026.
"We'll just go back to on-site interviews." If you're hiring globally, this kills your pipeline. The best Series A–C scale-ups across the UK, Nordics, and DACH region compete for senior engineers based in Warsaw, Lisbon, Tallinn, Bucharest, and Kyiv. "Fly every candidate in" means "hire 30% of the candidates you would have hired, at 3x the cost, six weeks later." A tax most scale-ups cannot afford.
"We'll just use harder problems." Frontier models in 2026 solve every standard LeetCode-tier problem in seconds, and a competent candidate only needs to relay the answer. Making the problem harder makes it harder for honest candidates too, and does nothing to the cheaters. Worse, it biases your funnel toward candidates who are good at looking at model output and repeating it convincingly.
"We'll watch more carefully for tells." Engineering hiring managers are not trained interrogators, and the modern cheating stack is specifically designed to eliminate tells. Eye movement looks normal because the overlay is on-screen. Typing cadence looks normal because the candidate is reading. Audio latency is under half a second. You cannot out-observe this.
"We'll add a take-home." Take-homes in 2026 are a pure AI signal. Unless you are actively measuring how the candidate uses AI — which can be a legitimate signal, more on that below — the take-home tells you only that the candidate has access to Claude, Cursor, or ChatGPT.
The problem is not that any single countermeasure is worthless. It's that they need to be stacked, and they need to be paired with a fundamentally different theory of what an interview is measuring.
A new theory of the interview
The assumption of the pre-2022 interview was: if the candidate produces the artifact (working code, a correct answer, a reasonable system diagram), they can do the job. That assumption is dead. The artifact is trivial to generate.
The new assumption has to be: the interview measures the candidate's ability to reason about the artifact in real time, under conditions where the artifact alone is not sufficient evidence. In practice, that shifts the interview in three directions at once.
First, you stop measuring the answer and start measuring the derivation. A candidate who pastes in the right answer and cannot then modify it under a new constraint, or cannot explain why one line was written the way it was, is giving you a signal — the absence of reasoning is the signal.
Second, you make the interview adversarial to pre-prepared context. Unique, proprietary, up-to-the-minute problems beat famous problems. A bug in a snippet of your own codebase (suitably anonymized) beats a stock LeetCode prompt. A broken PR with three subtle issues beats "implement LRU cache."
Third, you stop fighting AI and start measuring AI fluency. A senior engineer in 2026 should use AI. The question is whether they use it well — whether they can prompt effectively, recognize a wrong answer, push back on a confidently stated hallucination, and integrate model output into a larger piece of engineering judgment. This is a skill. It can be tested directly, and candidates who cheat on the "no AI" rounds often fail catastrophically on the "AI-allowed" rounds because they cannot drive the tool themselves.
A defensible 2026 interview loop
Here is a structure we have seen work across scale-ups hiring senior engineers, and that we build around at Recruo.
Stage 1: Asynchronous AI-aware screen. Give the candidate a realistic, scoped coding task and tell them explicitly that they may use any AI tool they like, but must submit a short recorded walkthrough explaining their decisions. You are measuring judgment under AI, not memorized syntax. This also filters out candidates who only function with an over-the-shoulder co-pilot, because the walkthrough exposes them.
Stage 2: Live secure technical interview. This is the round that has to be locked down. The candidate works in a controlled environment: a browser-based IDE, camera on, second-device detection in place, and no ability for external tools to read or write to the workspace. The problem is bespoke — ideally drawn from your own codebase — and the interviewer pushes the candidate off-script with follow-up constraints that no pre-generated answer will cover. This is where identity verification and anti-fraud controls pay for themselves.
Stage 3: Collaborative system design. System design is the round that is hardest to fake in real time, because the model cannot easily follow a two-way whiteboard conversation that keeps branching. Make this round genuinely conversational. Introduce business constraints mid-discussion ("the CFO just told us we can't use any managed Kafka for budget reasons — what changes?"). Watch whether the candidate thinks or recites.
Stage 4: Bar-raiser with a proprietary artifact. Final round: the candidate reviews a real piece of code or architecture from your product, with the worst bits already extracted and with a story. They walk you through what they would do, what they would keep, and what they would tear up. You cannot prepare for this in advance, and a model cannot meaningfully help a candidate who does not understand what they are looking at.
Across all four stages, the key pattern is the same: the interview is no longer a test of whether the candidate can produce an artifact. It is a test of whether they own the reasoning behind it.
The identity problem, and why it's separate
Everything above assumes the person on the call is the person who will show up on Monday. That assumption is no longer free.
At the senior end of the market, especially in remote-first scale-ups, we now recommend that every engineering hiring loop include an explicit identity-verification step before an offer is signed. This does not have to be adversarial. It can be as simple as a live, camera-on working session with the person's government ID visible, a short "tell me about your last project" conversation with pointed follow-ups about specific dates, tools, and teammates, and cross-reference against public footprint (GitHub commits, conference talks, LinkedIn connections with real people at prior employers).
If that sounds paranoid, it is the appropriate level of paranoia for 2026. Proxy interviewing has moved from a rumor into a documented, commercialized practice. Any CTO hiring remotely into senior positions who does not have an identity-verification layer is running an open door.
What this looks like at Recruo
We built our process specifically for the post-LLM world. Every candidate we present goes through AI-aware screening, secure live technical rounds conducted inside Recruo Secure Browser (which blocks second-device and overlay-based cheating by design), human review of soft skills, and a dedicated AI Skills Validation step that scores how well the candidate actually uses AI in realistic engineering work. The result is that when we send you five candidates, you spend your time choosing between them — not trying to figure out which one is real.
Our clients see an 80% pass rate on their own final-round interviews, an average time-to-hire of five days, and placement fees at 15% — roughly 32% below the standard agency rate. More importantly, they stop making the bad hires that the 2022 interview stack quietly started letting through.
What to do this quarter
If you're a CTO or VP of Engineering reading this, three things worth doing before the end of the quarter:
Audit your last four engineering hires who didn't work out. How many of them interviewed suspiciously well? That pattern is your baseline fraud rate, and it's almost certainly higher than you think.
Rewrite your technical loop around reasoning, not artifacts. Even a small change — replacing one stock problem with a bug from your own codebase — meaningfully raises the floor.
Decide whether you want to build anti-fraud tooling in-house or work with a partner whose entire pipeline is designed around this problem. The first is expensive, slow, and rarely a core competency. The second is what we do.
→ Ready to stop hiring the wrong person? Book a 20-minute call and we'll walk you through our interview loop, Recruo Secure Browser, and how we'd run your next senior engineer search.
Related reading: The True Cost of a Bad Senior Engineering Hire in 2026 · How Recruo's 6-Step Process Works