In early 2025, a Columbia University student named Chungin "Roy" Lee built a tool called Interview Coder: an invisible overlay that sits on top of your screen, reads the interview question in real time, and feeds you an AI-generated answer that's visible only to you - not to the interviewer's screen share. He used it to pass a software engineering interview at Amazon. Then he posted about it publicly. Amazon rescinded the offer within hours, and Columbia suspended him. Months later, his Harvard admission offer was separately rescinded after it emerged he hadn't disclosed a high school suspension. He then rebranded the tool as Cluely, raised a $5.3 million seed round from Abstract Ventures and Susa Ventures, and closed a $15 million Series A led by Andreessen Horowitz two months after that.
That's the arms race in one story: the tools work, the consequences are real, and the market for them is growing anyway.
What the tools actually do
Interview Coder and Cluely aren't the only players, but they're the ones that made the cheating conversation mainstream. The mechanism is consistent across this category: an invisible screen overlay listens to the interviewer's audio or reads the on-screen question, generates a response with an LLM, and displays it in a layer that standard screen-sharing software doesn't capture - typically within 1-2 seconds of the question being asked.
That last detail is the important one. Older cheating detection (tab-switch alerts, browser lockdown, webcam monitoring) all assumes the cheat lives somewhere the proctoring software can see: a second tab, a second monitor, a phone in frame. Overlay tools are specifically engineered to defeat exactly that assumption by operating beneath the screen-share layer entirely.
The scale isn't a fringe phenomenon anymore. Fabric, a company that analyzes interview sessions for cheating signals, found that 35% of candidates showed signs of cheating in late 2025 - more than double the rate from six months earlier - across a dataset the company later expanded to 19,368 analyzed interviews, where the figure came in at 38.5%. Separately, roughly 1 in 5 US professionals now admits to secretly using AI during a live interview when asked directly. The FBI has also flagged a related but distinct thread: state-sponsored actors, including North Korean IT operatives, using these same overlay techniques to get hired into Western companies under false identities.
The consequences are catching up to the tools
Roy Lee's case got attention because he made it public, but the pattern it revealed - discovery leads to fast, hard consequences - has become standard practice. Companies are increasingly writing explicit clauses into offer letters that allow rescission, or even termination after start, if interview cheating is discovered later. The Amazon case shows how fast that moves: hours, not weeks, between a public post and a pulled offer.
This is also part of why in-person final rounds have made such a strong comeback across the industry - a trend we cover in more depth in our piece on whether the coding interview is dead in 2026. The short version: an overlay tool that's invisible to screen-share is not invisible to a person sitting across the table watching you type. System design rounds are also creeping earlier into loops for the same reason - a live whiteboard conversation is much harder to run through a hidden tool than a solo coding prompt. Our system design interview cheatsheet covers the structure that format actually rewards.
How proctoring vendors are fighting back
The major coding-assessment platforms have all built countermeasures, with mixed and vendor-claimed effectiveness:
- HackerRank runs a Proctor Mode that combines screenshot analysis, plagiarism detection against a pattern database, and webcam anomaly detection, alongside a Secure Mode that enforces full-screen, blocks copy-paste, and flags tab switches. The company has claimed accuracy figures in the low-to-mid 90s percent for its combined ML and behavioral-analysis approach, though as with any vendor-reported detection metric, that number should be read as a claim rather than an independently audited result.
- CodeSignal uses what it calls a Suspicion Score, which analyzes submitted code for patterns consistent with GenAI output across its historical interview dataset, and separately tracks clipboard paste events - what was copied and from where - for human review.
- Karat, which runs technical interviews with live human engineers rather than automated assessments, takes a structurally different approach: its argument is that cheating is materially harder against an experienced live interviewer who can ask an unscripted follow-up and watch the code get written in real time, rather than against an automated grader evaluating a final submission.
The honest state of the technology: browser-based telemetry (tab switches, clipboard events, webcam checks) catches unsophisticated cheating, but purpose-built overlay tools are explicitly engineered to sit below that detection layer. No vendor claims to have fully solved this for asynchronous, unproctored formats, which is a meaningful part of why high-stakes technical rounds keep moving back to a human in the room.
What actually counts as cheating, versus permitted AI use
This is the part that gets flattened in most coverage, and it matters, because "using AI in an interview" is not automatically cheating in 2026 - a lot of companies now build interview formats that explicitly require it.
Several employers have started building interview formats around exactly this. Greenhouse - whose own applicant tracking system runs a large share of the industry's hiring - publishes a candidate-facing policy stating that candidates may be encouraged to use AI to research a topic or iterate on a coding exercise in certain interviews, but the final work needs to reflect the candidate's own synthesis and decisions, and candidates should expect to walk through their thinking, including how they used the tool. That's the same shape of format our coding interview piece found at AI-native labs like Anthropic and OpenAI: pair programming on real code, with AI tool use not just permitted but expected, and the interview measuring how well you drive the tool rather than whether you touched one at all.
The distinction that actually holds up across all of this: disclosed, in-format AI use that you can explain and defend is the new skill being tested. Undisclosed, hidden AI use designed to make it look like the answer came from you unaided is cheating, regardless of the company's general stance on AI. An overlay tool is built entirely around the second definition - hiding the AI's involvement is the product's core feature - which is why it stays disqualifying even at companies that are otherwise enthusiastic about candidates using AI well.
We cover the honest version of the second category - practicing with AI tools openly and well, the way Anthropic and OpenAI candidates are now explicitly expected to - in our coding interview format guide, and in what AI skills software engineers actually need in 2026.
What this means for you
If a company hasn't told you AI is permitted, assume it isn't, and don't run an overlay tool hoping it won't be discovered - the Roy Lee case is the object lesson in how fast and how publicly that goes wrong, and rescission clauses mean it can catch up with you even after you've started the job. If a company has explicitly built AI into the round, as more are doing, prepare to use it well and openly: know how to explain every line you produced, and expect to be asked to. Review your interview prep checklist either way - the fundamentals of a strong, defensible interview performance haven't changed, only the tools sitting on the table have.
There's also a quieter, more common risk than getting caught mid-interview: relying on a hidden tool means you never actually build the skill the interview was checking for in the first place. If the overlay answers the system design question for you today, you still have to defend that same knowledge in the on-site, the 90-day review, or the next interview loop at the next company - none of which come with an invisible assistant. The candidates doing well in the current environment aren't the ones with the best-hidden tool. They're the ones who can sit down with a person watching and produce the same quality of thinking unassisted, because that's the actual bar every one of these formats, proctored or not, is trying to measure.
Sources
- Meet the 21-year-old helping coders use AI to cheat in Google and other tech job interviews - CNBC
- Interview Cheating in 2026: Cluely, Interview Coder, and Why Traditional Interviews Are Failing - Fabric
- AI Tools Can Help Job Hunters Cheat on Interviews and Coding Tests - Bloomberg
- Prevent and detect cheating in recruiting - CodeSignal
- How do recruiters see how candidates used AI during their tests? - HackerRank
- AI in Interviews: Cheating or the New Normal? - Karat
- Guidelines for using AI in our interviewing process - Greenhouse
