What Link Guard does with your data

This covers the FraudLens AI — Link Guard Chrome extension only: what it sends to the FraudLens AI backend, what stays on your device, and what the backend keeps. It does not cover the separate FraudLens web app or Outlook add-in.

In short

The extension exists to warn you before you pay or hand over credentials on a site that looks like a fake Mauritius bank or telecom domain.
Nothing is sent anywhere until you install it and browse, or you explicitly ask it to check something.
It never reads form fields, passwords, or anything typed into a page — only rendered, visible text.
Verdicts are computed by the backend's deterministic rules; an AI model may enrich the signals but never decides the outcome.
We don't sell data, run ads, or use analytics trackers. The only network calls are to the FraudLens AI backend.

01 Scope

Link Guard is a Manifest V3 Chrome extension that calls the FraudLens AI backend for every check — it never scores a page or message locally. This policy describes exactly what leaves your browser, what the backend does with it, and what's kept afterward.

The extension does not contact any server other than the FraudLens AI backend it ships configured for. It has no ad network, no analytics SDK, and no third-party tracking script.

02 What's sent, and when

Every network call is triggered by browsing activity or an explicit action — nothing is sent silently in the background beyond the automatic domain check described first.

Trigger	What's sent	Where
You navigate to or switch to a tab (automatic)	The full address of that tab, including its path — not the page content	`POST /api/check-url`
You click "Scan This Page"	The page's visible, rendered text only, capped at 5,000 characters	`POST /api/analyze`
You right-click → "Check selected text"	The text you selected, capped at 400 characters	`POST /api/analyze`
You right-click a link → "Check link"	That link's URL	`POST /api/check-url`
You click "Security Report"	The page's URL, plus DOM and network-surface signals its own JavaScript can already see (scripts loaded, cookies' flags, headers observed on its own requests) — never your keystrokes or field values	`POST /api/analyze-site`
You click "Report this site"	The domain or sender string you're reporting	`POST /api/report`

Note on the automatic check: because per-tab lookalike detection needs the domain of every site you visit, that domain (as part of the full tab URL) is sent to the backend automatically as you browse — this is the extension's core function, not an optional analysis step. It is not stored beyond the short-lived cache described in §5.

03 What's never collected

Form fields, password managers' contents, or anything typed into an input, textarea, select, or contenteditable region — the page-scan extractor deliberately skips all of these.
Your browsing history beyond the currently active tab at the moment of a check.
Account credentials, cookies' values, or payment details.
Any identifier tying checks back to you personally — the extension has no account system and sets no tracking identifier.

04 How it's analyzed

The backend runs deterministic checks first — domain/lookalike matching against a registry of known Mauritius bank and telecom domains, identity consistency, payment-context and lexicon rules. These alone can produce a verdict.

For richer language understanding, the backend may also ask a language model to extract signals from the text: which manipulation tactics appear, and exact quotes as evidence. By default this runs against a self-hosted model; if that's temporarily unavailable, the backend can fall back to a hosted provider (currently Anthropic's API) to keep the service available. Either way, the model only proposes enum-coded signals and quotes — it never sets the risk score or the final verdict, which is always computed by versioned, deterministic rules.

Unlike the FraudLens web app's paste flow, which redacts identifiers (account numbers, phone numbers, email addresses) in your browser before anything is sent, the extension currently has no client-side redaction step: the text you scan or select is analyzed as-is. URLs, domains, and OTP-style codes are always kept unredacted everywhere, because lookalike and payment-context detection depends on them.

05 What's stored, and for how long

In your browser

Recent checks (chrome.storage.local): up to 15 entries of domain, timestamp, and result state only — never message text or page content. Cleared if you uninstall the extension or clear its site data.
Per-tab result cache (chrome.storage.session): a checked URL's result is reused for about 10 minutes so switching tabs doesn't re-trigger a check; cleared when the browser closes.
Security Reports (chrome.storage.session): your last 5 reports, so you can reopen the full-page view; gone when the browser closes.

On the backend

Message and page text sent for analysis is processed to produce a response and is not written to a database. Community "report" events are the exception: they're stored as a one-way fingerprint of the redacted text and an HMAC-SHA256 pseudonym of your IP address — the raw text and the raw IP are never stored. Report events are purged after 90 days; the audit trail of resulting risk adjustments is purged after 180 days.

06 Permissions

Permission	Why it's needed
`tabs`	Read the active tab's URL to run the automatic domain check and update the toolbar badge.
`activeTab` + `scripting`	Inject the page-text extractor only when you click "Scan This Page" or "Security Report" — never automatically.
`contextMenus`	Add the right-click "Check selected text" / "Check link" actions.
`storage`	Keep the recent-checks list and cached results described in §5.
Host permission for the backend origin	Allow the extension to call the FraudLens AI API — Manifest V3 blocks requests to origins not explicitly listed.

07 Third parties

The extension itself talks only to the FraudLens AI backend. That backend, in turn, may call:

A self-hosted language model it operates, by default — no external transfer.
A hosted LLM provider (currently Anthropic) only as a fallback when the self-hosted model is unavailable, to extract the same bounded signals described in §4.

We do not sell, rent, or share your data with data brokers or advertisers, and we don't run analytics or advertising trackers of any kind.

08 Your controls

Uninstalling the extension removes everything it stored locally (§5).
You can clear the recent-checks list from the extension's popup, or clear its site data from chrome://extensions.
"Scan This Page," "Security Report," and the context-menu checks only ever run when you trigger them.
If you've submitted a community report and want it removed, contact us with the domain/sender and approximate date — we can locate it by fingerprint even though we don't store the raw text.

09 Changes to this policy

If what the extension collects or how it's processed changes, this page will be updated and the "Effective" date above will change accordingly. Material changes will also be reflected in the extension's Chrome Web Store listing notes.

10 Contact

Questions or requests about this policy dhruvksbisht@gmail.com