Applies to the Chrome extension “Cybersecurity Browser Extension” (ID bihhhokkeemejefomgfkhfekibekjclf). Last updated 25 September 2026.
Cybersecurity Browser Extension is a research project from HELP University that detects phishing web pages and malicious links with machine-learning models. It is designed so that your browsing data stays on your device. This page explains exactly what the extension reads, what it stores, and what leaves your computer.
Changes on 25 September 2026: the Claude API key is kept in memory by default and only saved to disk if you choose Remember on this device.
Changes on 24 September 2026: contributing to model improvement is now your own opt-in switch (off by default) rather than an administrator setting; the random client identifier is no longer sent; flagged verdicts are now kept for 180 days; administrator sign-in is checked by our model service; the request logs for contributions are not kept.
The 22 September 2026 version replaced the one dated 11 April 2026. Changes: tabs and webNavigation permissions removed;
webRequest, identity, declarativeNetRequest and privacy added; ad blocking, Privacy Shield,
federated learning and on-device inference documented; knowledge-base editing restricted to administrators.
| Permission | Used for |
|---|---|
activeTab, scripting | read and analyse the page you asked to scan; outline or remove flagged links |
storage | your settings, results and history, stored locally |
cookies | the Cookies Analyzer view |
webRequest | observe response headers of the top-level page to derive 14 security signals (observation only; values never stored) |
declarativeNetRequest | Ads Remover and Privacy Shield rule lists; all off by default |
privacy | the optional WebRTC leak protection setting |
identity | Google sign-in for administrators only |
host permissions http://*/*, https://*/* | scanning, header observation and ad hiding can be needed on any site; the extension cannot know in advance which site is phishing |
Removed since the previous version: tabs and webNavigation.
| Data | Where | Purpose |
|---|---|---|
| Your settings (which protections are on, highlight style, allow-listed sites, chosen model) | Chrome extension storage | remember your choices |
| Last scan verdict, scan history, dashboard counters | Chrome extension storage | show results and statistics in the popup; you can clear them at any time |
| Reports you submit on the Report Phishing Website page (URL, label) | the local detection agent on your machine | on-device model training; the reports themselves are never uploaded |
| Wrong verdicts you flag on the Page Scan panel (three model scores, label — no URL) | the local detection agent on your machine | shown on the Federated Learning page; shared only if you turn on Contribute to model improvement |
| Your Claude API key (optional) | in memory until the browser closes; in Chrome extension storage on this computer only if you tick Remember on this device | only if you use the website-audit chat; sent only to api.anthropic.com |
| Administrator session token (administrators only) | Chrome session storage | deleted when the browser closes |
| Sites you allow ads on; Privacy Shield switches | Chrome extension storage | remember your exemptions and protections |
A small number of functions are restricted to administrators who sign in with Google: reviewing and publishing new model versions, and viewing or editing the chatbot's knowledge base. The sign-in is checked by our model service in Google Cloud, only to compare the account's e-mail address with a list of administrators; no Google profile data is requested or stored, and ordinary users never sign in. Training on your own device is not an administrator function — it is controlled by your own Contribute switch (section 3). The chatbot itself remains available to everyone.
| Destination | What | When |
|---|---|---|
Local detection agent (127.0.0.1:5000, on your own computer) | the page URL, the compact text representation, the 43 features and the 14 header signals | on every scan |
Our model service (Google Cloud, project browser-extension-509309) | requests for the list of current model files and time-limited download links; nothing about you or your pages | when the agent checks for model updates |
| Our model service — contributions only if you turn on Contribute to model improvement (off by default) | Flagged verdicts: the three model scores and the correct label. No URL, no page content. Training updates: after a training round on your device, the model update it produced (a few MB of numbers), the number of reports used, the model version it was trained from, and the size of the change. No URL, no page content. Neither carries a user, account or device identifier. | each time you flag a verdict; after each training round you start |
| api.anthropic.com | the text you type in the audit chat and, if you tick “include page context”, the page text | only if you have entered your own Claude API key and use the chat |
| Our model service (sign-in) | a Google OAuth token, verified with Google to confirm the account is on the administrator list | administrators only |
Never sent anywhere: raw page HTML, your browsing history, cookie values, header values, screenshots, keystrokes, or any personal identifier. The Ads Remover records nothing at all about which pages you visit or what it blocked; it keeps only a count of hidden elements for the open tab, in memory.
The developers operate no analytics, no crash reporting and no account system for users. The only data kept on our side are the contributions described in section 3, and only from people who switched contributing on:
| Data | Kept for |
|---|---|
| Flagged verdicts (three scores and a label) | 180 days, then deleted automatically |
| Training updates | 30 days, then deleted automatically |
Each is stored under a random name with only the day it arrived. We do not keep the request logs for these uploads, so your IP address is not stored with them (Google, which runs the servers, necessarily sees it while delivering the request). Because nothing links a contribution to a person, we cannot find or delete an individual's contributions on request; they expire as listed above.
A limitation you should know about. A training update is calculated from the reports on your device. It contains no URLs, but research has shown that a model update can sometimes reveal something about the data it was trained on. To reduce this, a training round requires at least 32 of your reports, and the size of every update is capped before it leaves your device. These measures reduce the risk; they do not remove it entirely. If that matters to you, leave contributing off — detection works the same either way.
We do not sell, share or transfer user data to anyone, and we do not use it for advertising, credit or lending purposes.
If the extension gains a capability that changes what data it reads or sends, this page is updated before that version is published, and the date at the top changes. The Chrome Web Store listing always links to the current version. Material changes are also noted in the extension's release notes on GitHub.
Questions about this policy: b2300174@helplive.edu.my (Faculty of Computing and Digital Technology, HELP University).