How do agentic AI browsers like Comet or Atlas get used to cheat on exams?
TL;DR: Agentic browsers such as Perplexity Comet or OpenAI Atlas can read what's on screen, fill in forms, and take actions on a candidate's behalf — turning an exam tab into something the browser itself can answer, not just look up. Perplexity's own CEO has publicly said he was troubled to see students using Comet this way. Proctoring needs to detect the browser process and automation behavior itself, not just block a ChatGPT tab.
Why this is different from copy-pasting into ChatGPT
Traditional AI cheating means switching tabs, pasting a question into a chatbot, and typing the answer back — a detectable tab-switch and typing pattern. An agentic browser can, in principle, read the question directly off the page and fill in the answer automatically, without the same obvious back-and-forth, since the browser itself is the agent.
Traditional copy-paste cheating vs agentic browser cheating
| Factor | Copy-paste to chatbot | Agentic browser |
|---|---|---|
| Tab switching required | Yes, usually detectable | Not necessarily |
| Typing pattern | Manual, human-paced | Can be near-instant or automated |
| Detection method | Tab-switch/clipboard monitoring | Process/extension detection, DOM automation signatures |
| Footprint | Visible in browser history | Can be less visible without deeper monitoring |
How to defend against it
- Enforce a locked-down exam browser that blocks unauthorized processes and extensions, not just specific URLs.
- Monitor for known agentic browser executables and browser-automation signatures running in the background.
- Watch for response timing that's too fast or too uniform for the question's difficulty.
- Combine with webcam/gaze monitoring — agentic tools still require the student to read or approve outputs.
- Update your blocklist regularly; new agentic browsers ship frequently in 2026.
FAQ
Are Comet and Atlas built for cheating? No, they're general-purpose productivity browsers — but their form-filling and screen-reading abilities are being repurposed for exam cheating.
Can lockdown browsers stop agentic browsers? A properly configured lockdown browser blocks unauthorized apps and processes from running during the exam, which covers this case if it's kept current.
Is this a bigger risk than regular ChatGPT cheating? It's a newer, less-studied risk because the automation reduces the usual behavioral tells of manual copy-paste cheating.
By Pinal Dave Last updated: 2026-07-26