A silent vulnerability exists in unprotected online exams that directly undermines the integrity of EASA aeronautical licences. It requires no screen switching, triggers no alarms, and cannot be detected by an exam invigilator. The risk is especially critical when students sit exams on personal, unaudited devices — with or without fullscreen lockdown.
The digitalisation of exams in Part-147 training organisations has brought undeniable advantages: flexibility, accessibility, and reduced administrative costs. But it has also opened a gap that demands immediate attention.
The problem: an AI inside the browser
Tools like HARPA AI are extensions installed directly into Google Chrome and similar browsers. Unlike opening a new tab or window — actions that a detection system can log — these extensions operate as a sidebar panel within the same browser window.
This means a student can be looking at the exam screen while simultaneously running an AI that reads the content of that very page and generates answers, without ever leaving it.
💡 HARPA AI can automatically read the content of the active page. In an exam context, this includes the questions, answer options, and any text visible on screen.
HARPA AI is just the tip of the iceberg
It would be a mistake to focus solely on HARPA AI as if it were the only risk. In reality, HARPA is simply the best-known example of an entire category of browser extensions that share exactly the same behaviour: they operate as a sidebar within the same browser, read the content of the active page, and generate AI responses without any need to switch windows.
Among the alternatives with an identical risk vector are tools such as Sider AI, Monica AI, Merlin AI, Voilà AI Assistant, Chippy, and Nanobrowser — all freely available on the Chrome Web Store. Banning HARPA AI by name in an exam policy solves absolutely nothing. A student can use any of these alternatives with the same result. The only effective solution is technical and operates at the device level, not the tool level.
Why fullscreen is not enough
Many online exam platforms enforce fullscreen mode and close the exam if the student leaves that window. It is a reasonable measure, but it has an important technical limitation.
Fullscreen mode hides the Chrome toolbar and tab bar, but it does not disable installed extensions. These remain active, continue reading page content, and continue responding. The detection system registers that the student never left the exam screen. And technically, it is correct. Neither did the AI.
But there is an even more exposed scenario: systems that do not implement fullscreen lockdown at all. In these cases, there is no technical or visual barrier whatsoever. The student can interact freely with the AI extension without any restriction, in complete comfort, with zero risk of triggering any alert.
🔴 An exam taken in standard Chrome, without mandatory fullscreen, on a personal unaudited device, represents maximum exposure. There is no technical barrier, no extension restriction, and no possible detection. It is the digital equivalent of sitting an exam with an open book on the desk.
The decisive factor: whose device is it?
The level of risk does not depend solely on the type of device or screen lockdown — it depends on who manages it.
A student accessing the exam from their own personal computer may have had HARPA AI installed for weeks or months before the exam. No fullscreen-based detection system can identify this. In contrast, a device managed by the Part-147 organisation through MDM policies can have the installation of unauthorised extensions blocked entirely, eliminating this risk vector before the student even sits down.
Full risk matrix
| Device | No fullscreen personal device |
Fullscreen personal device |
Organisation- managed device |
|---|---|---|---|
| 💻 PC / Laptop + Chrome | Maximum critical | Very high | Low |
| 📱 Android Tablet | Critical | Medium | Low–negligible |
| 🍎 iPad / iPhone | No risk | No risk | No risk |
| 🛡 Lockdown Browser (SEB) | N/A | No risk | No risk |
The matrix reading is clear: risk is highest on personal computers with no screen protection at all. Fullscreen mode slightly reduces the convenience of cheating but does not prevent it technically. Only the combination of a managed device or a lockdown browser eliminates the risk effectively.
✅ Organisation-managed Android tablets represent an option with a strong security profile and contained costs. Through MDM policies, the organisation can block the installation of alternative browsers such as Kiwi Browser — the only one that supports desktop extensions on Android — effectively eliminating the only meaningful risk vector on these devices.
Recommendations for Part-147 organisations
Immediately discontinue exam environments without fullscreen on personal devices. This is the maximum critical risk scenario. If device control is not possible, Safe Exam Browser must be mandatory.
Do not allow exams from unaudited personal devices. Fullscreen provides a minimal barrier, but HARPA AI and its alternatives remain fully operational. Only device-level control eliminates the real risk.
Consider organisation-managed Android tablets as a secure and cost-effective alternative. With correctly configured MDM policies, they eliminate the most relevant risk vectors.
Deploy Safe Exam Browser (SEB) or equivalent whenever exams must be taken on computers of any kind.
Apply MDM policies across all exam devices. Block unauthorised extensions and alternative browsers before the student arrives at the exam.
Do not rely on visual supervision by an invigilator. This type of fraud produces no visible sign. Human supervision alone is insufficient in any of the risk scenarios described.
Exam integrity is the foundation of trust in aeronautical licences. An EASA Part-66 licence obtained with undisclosed AI assistance does not only put aviation safety at risk — it puts at risk every person who depends on the technicians who hold it.
This article aims to inform the sector about an identified risk vector, so that training organisations can adopt appropriate preventive measures. It does not describe methods for actively using these tools.