Platform by Platform: What Each Coding Proctor Actually Flags

On 20 Aug 2026 I opened six pages on this question in six tabs: HackerRank’s July and April 2026 release notes, CodeSignal’s fraud page, CodeSignal’s Suspicion Score article, Codility’s proctoring article, and Interview Coder’s homepage. Four loaded. Two returned a 403 to an automated fetch, which I’ll flag rather than paper over. The finding worth the hour is that HackerRank and CodeSignal are now looking for close to opposite things, while most write-ups on this still treat the two as one platform.

They aren’t one platform. HackerRank spent this year adding sensors. CodeSignal spent it tuning a classifier. That difference is what decides which AI tools work on HackerRank and CodeSignal, and it explains why the same tool can be invisible on one and obvious on the other.

HackerRank now watches the room, not only the code

HackerRank’s July 2026 release notes are unusually plain about what got upgraded. Screenshot Analysis, they write, “Detects a broader range of suspicious applications, including newer AI tools, during Proctor Mode assessments.” Read that again. It is a vendor saying out loud that the screenshot pipeline is being tuned against the current generation of overlay apps, which is a very different thing from generic anti-cheat marketing.

Object Detection got the same treatment. It “More accurately identifies and flags mobile phones and tablets in the candidate’s webcam feed, including when devices are only partially visible in the frame.” A phone half under the desk edge counts now. The April 2026 notes had already introduced the feature, describing it as flagging “mobile phones and tablets in the candidate’s webcam feed during an assessment”, so this is the second pass on the same signal in four months.

Two more deserve a slow read. Conversation Detection “Recognizes a broader range of communication patterns in the code editor, including regional languages written in English, giving you greater visibility into potential external communication during assessments.” The April notes give the mechanic behind it: “Code is now analyzed during assessments to detect when candidates type and delete messages that may indicate communication with others.” Typing a question into the editor, reading a reply, then deleting the whole thing is itself the logged event. Deleting doesn’t undo it.

Then there’s a gaze signal, listed as limited availability, that flags a candidate repeatedly looking away from the screen and then resuming typing. Plagiarism scoring also moved to language-specific thresholds, which HackerRank says “account for the natural differences in code length across programming languages”. Stack those together and Proctor Mode is running six jobs in parallel: screenshots, webcam images, object recognition, editor-content analysis, similarity scoring, and eye movement. The April notes add that integrity reports now carry “clearer explanations and supporting evidence, making it easier to understand why a candidate was flagged”, which cuts both ways. Recruiters get a readable reason. Candidates get a readable reason attached to their name.

Here is the caveat that most of this discussion skips. None of the six runs on a plain unproctored assessment link. Proctor Mode and Desktop App Mode are per-assessment settings a hiring team switches on, and plenty of teams don’t, because camera requirements hurt completion rates. I don’t have a reliable number for how often it’s enabled, and I’d distrust anyone who quotes you one.

CodeSignal’s Suspicion Score is a similarity engine wearing a proctor’s coat

CodeSignal’s cheating and fraud page describes the Suspicion Score as something that will “Analyze multiple fraud signals, including solution similarity, telemetry, and copy-paste activity.” Its AI line is separate: “Identify patterns that may indicate AI-assisted behavior, which are often missed by basic plagiarism checks.” Proctoring is blunter than anything on HackerRank’s pages. CodeSignal says it will “Record video, audio, and screen activity for the entire assessment session”, and there’s identity verification using “a valid, government-issued photo ID.”

Notice the ordering. Similarity comes first, telemetry second, copy-paste third, behaviour last. The Suspicion Score knowledge base article returned a 403 to my fetcher on 20 Aug 2026, so I’m working from the product page wording instead of pretending I read the internals.

That ordering has a practical consequence people underrate. An overlay that never touches your clipboard still produces a solution, and the solution is the artefact. Clean, idiomatic, optimal on the first attempt, with no wrong turns in the edit history, and shaped like something the corpus has seen thousands of times before: that’s a similarity and telemetry story, not a screen-recording story. Retyping generated code by hand doesn’t launder it. Screen invisibility is close to irrelevant against a signal that never looks at your screen.

Codility and CoderPad: quieter, and partly unverified

Both Codility’s behavioural-events proctoring article and its AI-cheating post returned 403s to my fetcher on 20 Aug 2026, so I’m keeping this section thin on purpose instead of guessing. From the public product pages, Codility’s two mechanisms are Similarity Check and per-test behavioural event detection, including copy-paste tracking that lands in a timeline on the candidate report. CoderPad is a different animal: live pairing, so the detector is a human being on the call watching your cursor stall for nine seconds before perfect code appears. That one is the least automated and the hardest to beat.

So which AI tools work on HackerRank and CodeSignal?

No AI tool is reliably undetectable on both. On HackerRank with Proctor Mode enabled, screen-reading tools are the explicit target, because screenshot analysis is being tuned for “newer AI tools” by the vendor’s own description. On CodeSignal, screen invisibility barely helps, since the Suspicion Score reads your solution and your telemetry rather than your monitor.

Audio-only help lasts longer than screen overlays on both stacks, which is the honest version of the answer. It’s also a narrower window than the marketing suggests. Interview Coder’s homepage on 20 Aug 2026 claimed “20+ of the best undetectability features in the planet, tested every single day against every single interview software for 100% undetectability”, with per-platform rows for HackerRank, CoderPad and Codility, each marked Undetectable and stamped with a last-updated time measured in hours. Pricing on the same page that day: Free Download at $0, Monthly Pro at $299 a month marked down from $499, and Lifetime Pro at $799 one-time from $1,598. I can’t test a 100% claim and neither can the sites repeating it. What I’d say is that a 100% figure and a vendor shipping gaze tracking cannot both be right for long.

The part I’d argue about

My opinion, and it might be wrong: the detection arms race on coding assessments is already decided, and the platforms won it by moving the fight to the solution rather than the screen. You can hide a window. Hiding an edit history that looks like thinking is a much harder engineering problem, and nobody selling a $299 subscription has solved it. If you want the longer version of that argument, we wrote it up in how interview AI detection works and in can interviewers detect an interview copilot.

Where AI assistance still earns its keep is the spoken round, which nobody has automated a detector for, because the artefact is a conversation and not a diff. Our own figures, put where you can check them: LastRound AI’s pricing page on 20 Aug 2026 lists Free at $0 with 15 credits a month, Starter at $19 with 120 credits, Professional at $49 with 390, and Ultimate at $99 with 990, and those monthly credits reset each month. The AI interview copilot is built for live behavioural and system-design conversations. On a Proctor Mode coding assessment with the camera on, I’d close it, and I’d rather tell you that than sell you a claim the release notes contradict.

What I’d do with 48 hours before one of these

• Find out which mode you’re getting. The invite email usually says whether a webcam and the desktop app are required. Proctored and unproctored are different exams.
• Practise producing a messy edit history on purpose: a naive solution first, then the optimisation. Telemetry that looks like thinking is the cheapest defence there is, and it’s also just how people solve problems.
• Keep your phone in another room for anything with a camera, given how specific HackerRank’s object detection wording has become.
• Don’t type questions into the code editor. Ever. That’s the single logged behaviour both April and July release notes call out by name.
• If the round is live on CoderPad, talk continuously. Silence is the tell, more than any tool.

The framing that gets people flagged isn’t tool choice, it’s assuming one answer covers every platform. Check the mode, read what the vendor published this quarter, and treat “undetectable” as a marketing word with a shelf life of about a release cycle. For the broader picture on where proctoring is heading, there’s AI proctoring in 2026. What would you do if the gaze signal shipped to everyone next month?