DomainGuard — Privacy Policy

Effective: 2026-07-25

Applies to: the DomainGuard browser extension (v2.1.0), native endpoint agent (v2.6.8), mobile companion app (iOS, v1.0.0), and dashboard backend (release v2.7.3).

DomainGuard is an enterprise AI-governance and data-loss-prevention (DLP) tool distributed by your organization's IT or security team. It is not a consumer product. This policy describes what the software collects, where that data goes, and how long it is kept.

If you are an end user reading this because DomainGuard is installed on your work device: your employer (the "Organization") operates the backend and dashboard and is the data controller. DomainGuard is a product of Network Security Cloud Analytics Private Limited (NSCA). Your data is processed only by the Organization's own backend — it is not shared with any third-party AI vendor.

A defining design property: sensitive content is classified on your device — the allow / warn / redact / block *decision* your Organization's policy applies to AI tools is made locally, and raw content is never sent to any third-party AI vendor to be scored. Because this is monitoring software on company-owned, consented endpoints, your Organization can also configure the console to record what a user attempted to send so an administrator can investigate it. What is recorded depends on the configured capture mode (§1): in the default capture mode this includes the matched value plus surrounding context; in redact mode the matched value is masked on the device first. Either way, that captured content is visible in the console only to designated administrators (§5) — analysts and auditors see the event, its data class, and the verdict, never the raw value.

1. What we collect

Baseline browser + endpoint visibility (when installed)

• URL and domain of pages visited (the address-bar value), page title, timestamp, and dwell time
• Event type — e.g. visit, clipboard_paste, file_upload, AI prompt/response
• Automation flag — whether the page shows Selenium / Puppeteer / Playwright signals
• Machine identity — a stable machine ID (from the agent, otherwise a generated UUID), hostname, OS, browser name/version, and — on agent-managed machines — the signed-in username
• Versions — extension and agent version

AI-firewall content (only when your Organization enables the AI firewall / DLP)

Content is scanned on the endpoint by a 63-pattern secret/PII detector plus an on-device classifier, and tagged by regulated data class (PII, PHI, PAN, secret, source-code, financial, credential). What is *recorded to the console for administrator review* depends on your Organization's capture (the default) vs redact mode — capture keeps the matched value with surrounding context so an admin can judge intent; redact masks the matched value on the device before it leaves. Even with no sensitive match, a short excerpt (up to ~300 characters) of a paste or AI prompt may be recorded so the activity is reviewable. All such recorded content is admin-only (§5). Note: the capture-vs-redact choice governs the browser extension; on the endpoint agent's AI-proxy and CLI-tool-hook paths (e.g. Claude Code, Cursor), matched values are always masked at the source ([DG_REDACTED:…]) before anything is sent — the raw value never leaves the device on that path:

• Clipboard pastes that match a sensitive pattern — a masked or context-limited preview, never the full clipboard
• AI prompts and responses on recognized AI-chat sites — a preview run through the same DLP
• Files and images attached to an AI prompt — text is extracted from the file (zip, docx, xlsx, pptx, pdf, csv) and images are OCR'd, all on the device, before upload; only a redacted excerpt and the data-class verdict are recorded. Unreadable attachments (encrypted, unknown, oversized) are flagged, not sent as clean. The raw file bytes never leave your device.
• Cert-pinned AI coding assistants (e.g. Claude Code, Cursor, Codex): where your Organization enables CLI hooks, the typed prompt is inspected locally and — only when flagged — a redacted excerpt is recorded
• AI-bound network destinations — the destination host and byte counts (metadata), even where content cannot be read
• An excerpt and data-class tag on every flagged event — the matched value plus surrounding context in capture mode, or the value masked on-device in redact mode

AI-tool inventory (agent, when enabled)

The agent can inventory AI tooling on the machine so admins have a roster:

• AI tools / apps in use (process name, install path, vendor, last seen) — discovery identifies that a tool exists; it does not read prompts or code inside it
• MCP tool calls (server, tool, argument shape, verdict) when MCP governance is enabled
• Installed skills / editor rules (SKILL.md playbooks, Cursor rules, Codex docs) — only the frontmatter and a content hash are reported, plus a prompt-injection scan verdict; the skill body itself is never transmitted
• Device posture (disk / firewall / AV / patch signals)

Mobile companion app (iOS)

The DomainGuard mobile app (for admins and security teams) is a companion to the web console. It collects only what it needs to sign you in and notify you:

• Your email address (to sign in) and a random device identifier (to bind your session and deliver push)
• A push notification token (only if you allow notifications)

Your session token is stored in the iOS Keychain, and Face ID / biometric unlock is performed by iOS locally — the app never receives your biometric data. In the app you view only your own Organization's (tenant-scoped) security activity, which is the Organization's data governed by the sections above. The app talks only to your Organization's own backend, contains no analytics, advertising, or third-party tracking SDKs, does no tracking, and never sells or shares your data. Accounts are created by your Organization's administrator — to delete yours, contact your administrator or use the web console.

What we explicitly do NOT collect

• Login-form fields or passwords typed into ordinary websites
• Full page bodies, or the content of non-AI sites
• Cookies or session tokens
• Every keystroke — it is not a keylogger
• Browsing in Incognito / Private windows, unless an admin explicitly enables it
• Pages whose URL begins with chrome://, edge://, about:, file://, data:, or blob:

We do not screenshot your device; images you *attach to an AI prompt* are OCR'd on-device as described above.

2. How it is transmitted

• All data travels over HTTPS/TLS.
• The extension and agent authenticate with a tenant-scoped API key and, where provisioned, a per-device credential — sent only to your Organization's configured backend, never exposed to web pages.
• The extension talks to a local helper agent on loopback (127.0.0.1) only, to read machine identity — no browsing data is sent to the local agent over any external network.

3. Where it is stored

• Event data is stored in a PostgreSQL database operated by your Organization. The managed deployment runs on Railway (currently a United States region); your Organization may self-host in any region, which is how storage residency is controlled today.
• High-value secrets the backend holds (compliance signing keys, SSO and SIEM credentials, MFA seeds) are AES-256-GCM encrypted at rest with a dedicated key. Bulk event data relies on the database platform's disk encryption.
• On the agent's AI-proxy and CLI-hook paths, content is classified on the endpoint; only decision metadata and — where your Organization enables it — redacted excerpts are stored server-side.

DomainGuard data stays within your tenant's backend; it is not shared with any third-party AI vendor.

4. How long it is kept

• A configurable retention window — default 60 days for AI-activity and event data, which your Organization may set to 7–90 days. (A separate sweep for the web-events table honors DATA_RETENTION_DAYS, default 90.) Older records are deleted by an internal sweeper; the tamper-evident audit chain is retained for compliance evidence. Check with your IT/security team for the policy in effect.

5. Who has access

• Organization administrators — authenticated dashboard and mobile users see events for machines in their own tenant only; every query is scoped to the tenant, and there is no path that returns another tenant's data.
• Raw captured content is administrator-only. The matched sensitive values and prompt/response excerpts are returned by the API only to users with the administrator role. Analysts, viewers, and auditors see the event, its data class, host, and verdict, but the content field reads *"Restricted — visible to admins only"* — enforced centrally on every AI activity endpoint, not per-page.
• Webhook / SIEM subscribers — if configured, summarized event metadata may be delivered to external systems your Organization controls (over HTTPS; webhook payloads are HMAC-signed).
• No third parties — DomainGuard does not share, sell, or rent collected data. There are no advertising trackers, analytics SDKs, or marketing pixels in the extension or the mobile app.

6. Your rights

Contact your Organization's IT or security team to see what data has been collected from your machine, request correction or deletion, understand what policies are being enforced, or object to processing. DomainGuard has no direct relationship with end users; your rights are administered by the Organization that deployed the software. Mobile-app and web accounts are created and managed by your Organization's administrator — to deactivate or delete an account, contact your administrator or use the web console.

Compliance evidence + tamper-evident audit

Your Organization can generate Ed25519-signed compliance evidence bundles (audit-log segments, policy snapshots, machine inventory, detection summaries, and per-framework coverage for SOC 2, CMMC, ISO 27001, NIST CSF / 800-53, OWASP LLM & Agentic Top-N, MITRE ATLAS, and the EU AI Act). Every administrator action is recorded in a SHA-256 hash-chained audit log; the public verification key is published at /api/compliance/pubkey/:tenantId so external auditors can verify offline.

7. Security

• TLS in transit everywhere
• bcrypt password hashing; SHA-256 / HMAC hashed API / enrollment / device / refresh tokens; AES-256-GCM for high-value secrets at rest
• Per-device credentials so a single endpoint can be revoked; optional strict device-auth
• Per-tenant data isolation enforced on every query
• JWT session auth with server-side revocation; rate limiting and account lockout on authentication; optional two-factor authentication (TOTP) for admin accounts
• The local helper agent listens only on loopback (127.0.0.1) and accepts no external connections

8. Consent & disclosure

DomainGuard is workforce-monitoring software. Where it is deployed via managed install (GPO / Intune / Google Admin Console), the deployment is the disclosure pathway — your Organization is responsible for notifying employees that endpoint and AI-usage monitoring is in effect, and for complying with applicable law (including data-protection, electronic-communications, and works-council requirements in your jurisdiction).

9. Changes

We may update this policy. The Effective date at the top reflects the most recent revision. Material changes are announced via the dashboard's release notes.

10. Contact

Questions about how your Organization uses DomainGuard: contact your Organization's IT or security team.

Questions about DomainGuard itself (the software): developers@networksecuritycloud.com.