Can You Take a Coding Assessment in an AI Browser?

The browser list on your assessment invitation was written to solve a rendering problem, not an integrity problem. It tells the platform your editor will load and your screen share will work. It was not designed to tell anyone whether the browser you opened has a model sitting beside the page, reading it.

That distinction now matters, because the browsers on those approved lists changed underneath them. Chrome ships Gemini. Edge ships Copilot with page access. And a separate class of AI-native browsers — ChatGPT Atlas, Perplexity Comet — presents itself to a website as ordinary Chrome. This article covers what an assessment platform actually verifies about your browser, what it does not, and what to do before your next proctored test. Facts are current as of 15 September 2026.

What the browser requirement actually checks

Assessment vendors publish short compatibility lists. HackerRank's candidate documentation names Google Chrome, Mozilla Firefox, and Edge as supported, and singles out one exclusion: "The Brave browser is not supported" (HackerRank browser recommendations for tests). CodeSignal recommends the "latest version of Chrome, Firefox and Edge" on a laptop or desktop, and states that mobile devices and tablets are not supported for proctored assessments (CodeSignal supported browsers).

Codility is the most specific, and the most revealing. Its list is version-gated: Chrome ≥ 131, Firefox ≥ 132, Opera ≥ 114, Safari ≥ 17, Chromium Edge ≥ 130. Screen sharing narrows to Chrome, Chromium Edge, and Opera, and Codility notes that "we currently only support Chrome for any recording and screenshooting related functionality." The enforcement it describes is an alert: you "receive an alert if our system detects that you are attempting to complete the assessment on an unsupported browser" (Codility supported browsers and devices).

Read those three together and the shape of the check is clear. It asks whether the rendering engine is recent enough and whether the screen-capture APIs will behave. A browser that reports a current Chromium version passes. Nothing in the published requirements asks what else that browser can do.

The approved browsers now have assistants built in

The more interesting consequence is not about exotic browsers. It is that "use Chrome" and "use Edge" stopped meaning "use a browser without an AI assistant."

Google announced an expanded Gemini integration in Chrome on 28 January 2026: a side-panel experience that works across tabs, connected apps, and an agentic "auto browse" capability described as handling multi-step tasks such as filling out online forms, initially rolling out in preview in the U.S. for Google AI Pro and Ultra subscribers (Google blog, "The new era of browsing"). Reporting in August 2026 described the Chrome integration completing its rollout to Android users in the U.S., reached from a toolbar shortcut beside the address bar, with a setting that controls whether the current page is automatically added as context to a prompt (9to5Google, 18 August 2026).

Microsoft documents the equivalent surface as an enterprise policy rather than a feature announcement. The CopilotPageContext policy controls whether Copilot in the Edge side pane can access page content, and Microsoft's documentation states the default outside the EU is enabled, with the EU default disabled and users able to change it either way (Microsoft Learn, CopilotPageContext). The policy applies to Microsoft Entra ID profiles.

Both features are reachable by a deliberate action — opening a panel, sending a prompt — rather than firing on their own while you read a problem statement. But a candidate who follows the instruction to use an approved browser is not thereby in an AI-free environment, and the published assessment rules generally do not acknowledge that the approved browser is now one of the AI surfaces. If you are trying to establish what is permitted rather than what is possible, verifying the AI rules before a technical interview is a separate exercise from checking the browser list, and availability has never been permission.

What an AI-native browser looks like from the page's side

ChatGPT Atlas launched on 21 October 2025 on macOS, with an agent mode in preview that performs actions in the browser on the user's behalf (OpenAI, Introducing ChatGPT Atlas). Perplexity's Comet occupies the same category.

The detail that matters for an assessment is how these identify themselves. Analysis by bot-detection vendor HUMAN Security reports that Atlas ships a default user agent matching Chrome on macOS with no dedicated Atlas token, and that its traffic is not distinguishable from Chrome-like human traffic at the protocol level; the firm argues that identifying agentic sessions requires behavioral classification rather than a signature (HUMAN Security on ChatGPT Atlas). That is a vendor's account from a company selling detection, so treat the capability claims accordingly — but the user-agent observation is straightforward and checkable.

The practical implication cuts in an uncomfortable direction. A version-gated compatibility list of the kind Codility publishes will admit a Chromium-derived AI browser without comment, because the thing it inspects is the version, not the product. Silence from the browser check is not clearance. It is the check answering a different question, the same way browser extension detection answers a narrower question than most candidates assume.

What has been demonstrated, and what is only claimed

Education has the clearer public record so far. Colorado State University's teaching centre published a walkthrough in November 2025 showing agentic browsers completing Canvas quizzes: Perplexity's browser reading questions and selecting options after a single instruction, and Atlas supplying answers in a sidebar, with a recorded 5/5 on a sample quiz. Its recommended mitigation was a lockdown browser (CSU TILT, "New AI browsers can complete Canvas quizzes").

On the other side, proctoring vendors market the countermeasure. Honorlock's page on blocking AI tools, updated 31 August 2026, says its BrowserGuard secures the browser so that tools like ChatGPT "can't be opened at all or are flagged for live proctor review," and describes targeting browser extensions and "invisible" assistants that render as transparent overlays (Honorlock, blocking AI tools). Proctortrack's own commentary is more candid about the gap, noting that because these tools work inside the browser environment, "traditional browser monitoring may not always detect their usage" (Proctortrack on AI browser extensions).

Those are vendor claims about vendor products, not independent test results, and none of them are hiring platforms. Treat them as evidence that the category is being actively defended, not as a measurement of what any particular assessment you are sent will do.

What to do before your next proctored assessment

The useful move is to stop reasoning about detection and start reasoning about what you have been told.

Check the instructions, not the browser list, for the AI rule. The browser requirement and the assistance policy are usually written by different teams and live in different documents. CodeSignal's General Coding Assessment rules, for instance, are a separate artifact from its browser support page.

Open the assessment in a clean profile of a browser you can account for. A fresh Chrome or Firefox profile with no extensions and no assistant panel signed in is easier to describe honestly afterwards than a daily-driver browser with six things attached to it. That matters less for detection than for the conversation that follows an integrity flag on a coding assessment, where what you can plausibly account for carries real weight.

Expect the assessment to move out of the browser entirely when the stakes rise. The direction of travel among hiring platforms is toward desktop lockdown applications rather than web pages, which sidesteps the whole question by not being a browser.

And if the rules do permit AI, use the tool the employer named. In an AI-permitted round the prompt transcript is typically part of the evaluation, and a side channel the grader cannot see is worth less than a visible one you can defend.

Where this leaves the question

You can technically load most assessments in an AI browser, and the compatibility check is unlikely to stop you. That is not the same as being allowed to, and it is not a statement about what a given proctoring stack records. The honest summary is narrower than either the vendor marketing or the panic: browser allowlists were not built to police AI, the approved browsers acquired assistants after the lists were written, and the platforms have mostly responded by moving high-stakes assessments into controlled desktop clients instead.

For candidates, the safe default is the boring one. Use a plain, current browser in a clean profile, read the assistance policy as a separate document from the technical requirements, and ask the recruiter when it is silent rather than inferring permission from the absence of a block.

Control is a desktop application for Windows and macOS rather than a browser extension or a web app, which is an architectural difference from anything that lives inside the page — not a claim about what any proctoring stack can observe. If you are still working out where the line sits for your own interviews, what platforms can and cannot infer about ChatGPT use is the more useful place to start than a browser requirements page.