Shadow AI Browser Extension

The extension is a lightweight Manifest V3 background service that passively observes when users visit known AI tools (ChatGPT, Claude, Gemini, Microsoft Copilot, GitHub Copilot, and ~60 others) and forwards anonymized sightings to AI Discovery in RAIC.

It does not read page content, intercept network requests, inject scripts, or capture keystrokes. Detection is based purely on the browser's tab URL and page title.

The extension requests the following permissions:

• tabs — to read the active tab's URL and title on navigation.
• storage — to buffer pending sightings and store the device token.
• alarms — to schedule the 5-minute sync interval.
• webRequest + <all_urls> — required only for optional API endpoint detection (see below). The listener exits immediately if your org has not enabled the feature.

Never captured: page DOM, form fields, prompts, responses, request/response bodies, request headers, query strings, cookies, localStorage of visited sites, or screen content.

In addition to detecting AI tabs users open, the extension can optionally detect background calls made by other apps to known AI APIs (e.g. api.openai.com,api.anthropic.com,generativelanguage.googleapis.com). This catches Shadow AI embedded inside internal tools, IDE plugins, and SaaS apps that wouldn't surface as a tab visit.

Opt-in per organization. Default is off. Org Admins toggle it under Admin Panel → Org Settings → Browser Extension. When off, the webRequest listener exits immediately and no API URLs are observed.

When on, each match captures only: the AI API hostname/path, the calling app's hostname (page domain), the page title, and — for v1.2+ — the specific AI model identifier (e.g. anthropic.claude-3-5-sonnet-20241022-v2:0,gpt-4o-2024-08-06) extracted from the URL path (AWS Bedrock, Azure OpenAI/Foundry, Google Vertex/Gemini) or, for OpenAI/Anthropic Direct, the single JSON model field from the request body — no other body fields are inspected. Same 10-minute dedup window applies per (source_domain, api_domain, model_id) tuple, so model switches surface as separate detections. Detections appear in AI Discovery with an API Call badge, a Detected Model column, and a Called by: attribution.

The extension syncs automatically on its own schedule via chrome.alarms. The Sync now button in the popup is a manual override that now refreshes the latest catalog, org config, and approved-list cache before flushing detections.

• Detection sync — every 30 minutes. Flushes the local detection buffer to the shadow-ai-ingest edge function.
• Config refresh — every 6 hours. Polls get-extension-config to refresh the org's API-detection toggle and tier eligibility.
• Catalog refresh — every 24 hours. Pulls the latest published AI domain catalog from get-ai-catalog.
• Approved-list refresh — every 24 hours. Polls get-org-approved-catalog for warn-enabled state and approved catalog entry IDs.
• On install / popup open. A silent reconnect runs immediately if the user is signed into RAIC in any tab.
• On RAIC sign-in. The web app fires a handshake event the extension picks up in real time.
• Manual "Sync now". Forces an immediate catalog/config/approved-list refresh and then flushes detections without waiting for the next scheduled refresh.

The device token persists in chrome.storage.local until it is revoked, expires from inactivity (90-day sliding window via last_seen_at), or the org's tier/status changes.

The shadow-ai-ingest edge function authenticates each device session, validates the payload, and writes results to discovered_ai_apps. All rows are tagged with org_id and protected by Row Level Security so each tenant only sees its own discoveries.

Device sessions can be revoked at any time from Admin → Integrations → Connected Devices.

The AI domain catalog is now dynamic. The bundledai-domains.js serves only as an offline fallback. On install and every 24 hours thereafter, the extension fetches the latest published catalog from the get-ai-catalogedge function and caches it in chrome.storage.local.

Platform Managers (Super Admins) maintain the catalog under Account Tools → AI Catalog. Edits are staged until a new version is explicitly published, and any prior version can be restored via Version History. All publishes and rollbacks are fully audited.

When an organization admin enables "Warn users about unapproved AI services" under Admin → Integrations → Browser Extension, the extension shows a non-blocking dialog whenever a user visits or invokes an AI service that is not on the organization's approved list.

A service is considered approved when it is mapped to a catalog entry on a registered AI system whose status is Approvedor Conditionally Approved. If a system is later disapproved, retired, or unmapped, it is automatically removed from the approved list.

The dialog reads: "This AI service is not on your organization's approved list. If you need it for business use, please submit a use case or AI system for review." It offers two actions: Submit for Review (opens the intake page) and Dismiss for 8 hours (per-host, per-organization suppression persisted in chrome.storage.local so the 8-hour cooldown survives browser restarts and new tabs on the same site — fixed in v1.5.4).

To minimize network traffic, the approved list is refreshed at most once every 24 hours per device using a server-side versioncounter — when the device's cached version matches, no payload is transferred. The list also force-refreshes when the AI catalog itself is bumped.