Is Dusky Detectable? An Honest Answer From the People Who Built It

Search “is [tool] detectable” for any AI interview tool and you’ll find two kinds of pages: attack posts written by a competitor, and vendors insisting their own tool is “100% undetectable.” We build Dusky, so read this knowing who wrote it. In return: mechanism, evidence, and the exact situations where the honest answer is “do not rely on it.”

Claims about other companies’ products were checked 2026-07-12 and are stated as of that date. Claims about Dusky were re-checked against the shipping build on 2026-08-09.

The short answer

On a normal Zoom, Google Meet, or Microsoft Teams call, with the app set up correctly: no. When you share your screen, Dusky’s window is removed by your operating system before the call software ever receives the pixels. The interviewer sees your desktop, your browser, your editor — no Dusky. It’s not a Zoom trick — the OS itself enforces the exclusion, and you can verify it in five minutes (steps below).

But “no” is not the whole answer. The protection does not apply everywhere: proctoring apps that read your process list, corporate machines you don’t control, a few unusual capture paths, and — the one nobody markets — your own behavior on camera. Each is covered below, because a stealth claim that hides its failure modes is not one you can act on.

How Dusky’s invisibility actually works

Dusky registers as a screen-share-excluded window at the operating system level — the system content protection API on macOS, the display affinity API on Windows. Video conferencing apps respect this flag and skip the window when sharing your screen. The full feature reference lives in our Stealth Mode documentation.

In plain English, drawn from a first-hand technical teardown of how these tools hide from Zoom and Microsoft’s API documentation: a native desktop app can ask the OS compositor — the layer that assembles everything on your screen — to drop its window from every capture surface, before any capturing app can read the pixels. Zoom cannot un-hide a window it never received.

This is also why some competing tools leak: a browser extension or web-tab overlay is drawn inside the page’s own pixels, so it’s captured like any other pixels — there is no separate OS window to exclude. Only a native app gets compositor-level exclusion. (More on specific tools in Is Cluely detectable?.)

On top of the API, a few Dusky design decisions:

• Always on. There is no stealth on/off switch to leave in the wrong position before a call.
• No Dock/Taskbar icon, no app-switcher entry while in stealth — even the corner hint pill shown while hidden is invisible to capture.
• The overlay is always capture-excluded — the dashboard is a separate case. Dusky has two windows: the floating overlay you use during a call, and an ordinary dashboard window for sign-in, settings, and billing. The overlay’s exclusion is unconditional. The dashboard is protected only while an audio session is active and not paused — and “session” here means audio capture and nothing else, so in a screenshot-only round with the microphone off the dashboard is never protected at all. Close it before you share your screen. (Stealth Mode docs)
• Recordings too: if a recording tool captures the window frame, the contents appear blank.

This works across Zoom (desktop, browser, any share mode), Google Meet, Teams, Webex, Slack Huddles, CoderPad, and most browser-based conferencing tools — details in the Zoom walkthrough.

Where Dusky does NOT protect you

This is the section competitor pages don’t write. Each item is a real limit, most already documented in our own docs and FAQ.

1. Setup failures you can prevent

• macOS Screen Recording permission is required. Without it, invisibility does not work — Dusky shows a “Core features are currently unavailable” banner. Granting it requires an app restart.
• Zoom’s capture mode. The default “Auto” usually works, but set Zoom → Settings → Share Screen → Advanced → Screen capture mode to “Advanced capture with window filtering” — independent testing confirms Zoom needs this mode for exclusion to behave reliably; old Zoom versions may need updating.
• The dashboard outside a live audio session — including while a session is paused, and throughout any screenshot-only round with the mic off — is a regular, capturable window. Don’t leave it open while you’re sharing your screen.

2. Machines you don’t control

Corporate MDM or group-policy settings can restrict the permissions Dusky needs; if the app can’t register them, screen-share exclusion may not work correctly. On an employer-managed machine, run the two-device check below first — and if the overlay shows up on the second screen, believe the second screen.

3. Capture paths that bypass the compositor

OS-level exclusion is the strongest mechanism available to a desktop app, but it is not physics-proof. The researcher who tested this class of protection found QuickTime’s screen recorder bypassed it entirely, and notes DirectX, kernel-level captures, or GPU frame grabs “might still see you.” Microsoft is explicit that display affinity is not a security guarantee — a phone camera pointed at your screen captures everything. On Windows, Dusky applies capture exclusion through the platform content-protection API rather than hard-gating a minimum OS version at install; the strongest “exclude from capture” behavior it can request requires Windows 10 version 2004 (build 19041) or later, so on an older build you should not assume the overlay is excluded without checking. OS updates can change capture behavior too — one reason “test before every real interview” — with a second device, not just the onboarding button — is standing advice, not fine print.

4. Proctored assessments that inspect your machine

A video call captures pixels. A proctoring app runs on your machine with permission to see what else is running. HackerRank’s Desktop App Mode docs say it “detects and closes any other programs that candidates try to open during the test, including invisible cheating tools such as Cluely and Ultracode.” Proctorio and Honorlock block such tools by preventing them from launching — not by spotting them in the share. Standalone detectors work the same way; as Proctaroo’s CEO told TechCrunch, “we can see running applications and ‘hidden’ background processes.”

Screen-share exclusion does nothing against process-list inspection. Dusky makes no invisibility claim for locked-down proctored assessments, and you should not rely on it in one.

5. You

The tell that catches most people isn’t software. Detection vendor Fabric — a hiring platform selling cheating detection, so treat its findings as vendor claims — describes a signature pattern: a consistent 3–5 second delay after every question regardless of difficulty, mechanical left-to-right reading eyes, and speech that grows more uniform as the interview goes on. A first-hand interviewer account describes the same tells: reading eyes, a delay before every answer, buzzword answers delivered like a script.

An invisible window doesn’t fix how you behave while reading it. That’s why Dusky gives you frameworks rather than word-for-word scripts and grounds answers in your own CV — you still have to sound like yourself, because you are the one thing no API can exclude from capture. Our Zoom how-to guide covers using it naturally.

Test it yourself: the five-minute check

The point of a reproducible mechanism is that you can reproduce it — and you should, because you are the only one who can.

Be clear about what the onboarding step does and does not do. Dusky walks you through an invisibility check during setup, but it is a set of instructions and a confirmation button: the app does not inspect anything, and it does not verify that you were actually hidden. It records that you pressed the button. It also only re-opens later if you skipped it the first time.

So the real test is the one you run yourself, with a second device. The thing to test is the overlay, during a live session, because that is the only combination a real interview puts you in:

1. On macOS, confirm Screen Recording permission is granted (no yellow banner); restart Dusky if you just granted it.
2. In Zoom, set Screen capture mode to “Advanced capture with window filtering.”
3. Close the dashboard — the ordinary window you sign in and change settings in. Two reasons, and both will otherwise invalidate your test: while the dashboard is on screen, Cmd/Ctrl+B raises and focuses it instead of toggling the overlay, and outside a live audio session the dashboard is deliberately a normal capturable window. What you want on screen is the floating overlay bar — press Cmd/Ctrl+B to bring it up if it isn’t already.
4. Join a test meeting from a second device or account, and from your interview machine share your entire screen.
5. Start a live session — Cmd/Ctrl+Shift+A, or the mic control on the overlay. Taking a screenshot does not start one; only audio capture does.
6. Trigger a response (Cmd/Ctrl+Enter). The second device should show everything on your desktop — except the Dusky overlay.
7. Press Cmd/Ctrl+B to hide the overlay and confirm the corner hint pill is invisible on the second device too.

If you tested with the dashboard open and outside a session, you tested the wrong window: you’d have seen a window that genuinely is captured, and learned nothing about the one that isn’t. Repeat before every real interview — not because the mechanism is flaky, but because your environment isn’t fixed: OS updates, Zoom updates, and IT policies all change underneath you.

Why we won’t say “100% undetectable”

As of July 2026, Interview Coder claims “100% undetectability” and “zero documented cases of users being detected or flagged when using this tool properly,” and Ultracode goes further: “Only ULTRACODE can GUARANTEE 100% Undetectability.” Meanwhile the HackerRank documentation quoted above names Ultracode — the tool offering the guarantee — as one it detects and closes. Absolute claims are how this category earned its trust problem.

And it isn’t one or two tools. Effectively every major rival markets absolute invisibility, and the people who build detection keep publishing the counter-evidence. HackerRank’s own engineering team documented catching Interview Coder: “clearly visible to interviewers through the HackerRank screenshare feature” on macOS, and on Windows any mouse interaction triggered an “Out of Interview” alert. HackerRank’s Desktop App Mode docs name Cluely by name among the “invisible cheating tools” it closes. Proctoring vendor Talview publishes dedicated “stop [tool]” detection pages for Final Round AI and Parakeet AI. LockedIn AI even publishes its own guide to sharing your screen without being caught — a tell in itself that “hidden” has conditions. And Parakeet’s own FAQ concedes the limit outright: “there is currently no way to change the ParakeetAI process name … in Activity Monitor or Task Manager.”

We quote that as evidence about the category, not as a stick to beat one rival with — because Dusky has exactly the same exposure and it would be dishonest to imply otherwise. Dusky is a native desktop app, and a native desktop app is a running process: on macOS it appears in Activity Monitor as Dusky, and anything that enumerates running programs can enumerate ours. We hide from capture, not from process inspection, and section 4 above says so without hedging. The honest reading of Parakeet’s admission is not “so use Dusky instead” — it’s that no desktop tool in this category is 100%. OS-level exclusion is strong on a plain video call and useless against a proctor reading your process list. Being the one vendor that doesn’t claim 100% is the position you can actually act on.

There’s a privacy version of the same honesty gap. Several rivals route your interview transcripts to their servers, and some go further: Final Round AI’s own privacy policy lists “test, train, and improve our AI models” as a use of your interview transcriptions and responses, and shares data with advertising and analytics partners. Dusky stores nothing server-side — transcription and AI responses pass through and are not retained by us — and your Interview Context (CV and job description) stays on your machine, encrypted with AES-256-GCM and wiped on sign-out.

Our claim is narrower on purpose: on a plain video call, correctly set up, screen share cannot see Dusky — mechanism, proof video, failure modes, and a test you can run. And one claim we’ll never make: that using Dusky is permitted in any particular interview. Dusky makes no representation about whether its use is allowed in any given context; that determination is yours alone.

Try it against your own setup

The free trial exists so you can run that check on your own machine, OS version, and Zoom install before an interview is on the line: 15 minutes of AI-assistance time (the clock only counts while Dusky is actively answering), every feature included, no credit card. Download at getdusky.app, run the test, and see pricing if it earns a place in your prep.

FAQ

Can my interviewer see Dusky when I share my screen on Zoom, Meet, or Teams?

No — set up correctly, the OS excludes Dusky’s window before the conferencing app receives the pixels, across Zoom’s share modes, Meet, Teams, Webex, and Slack Huddles. Run the two-device check before every real interview.

Does Dusky show up in screen recordings?

Dusky uses OS content protection: if a recording captures the window frame, the contents appear blank. One documented exception in this class of protection is QuickTime’s recorder. If your interview is recorded through an unusual local tool, test that exact path first.

Can HackerRank, Proctorio, or other proctoring software detect Dusky?

Proctoring apps don’t watch the screen share — they inspect your process list or block unapproved apps from launching. Screen-share exclusion offers no protection against that, and we claim none. Don’t rely on Dusky in locked-down proctored environments.

How do I test invisibility before a real interview?

Join a test call from a second device, share your screen, and confirm the Dusky overlay and its hint pill are absent on the other side. The onboarding step walks you through this, but it verifies nothing on its own — the second device is what tells you the answer. Full checklist above and in the Stealth Mode docs.