What the extension sees and what it sends

The extension sends Sorvek only the items described here, according to two features (SaaS visibility and AI analysis) and your organization's settings. For a message sent to an AI tool, it sends a 200-character preview masked on the device, which is kept, and up to 4,096 raw characters, which the server analyzes again and then discards without storing them. Everything is linked to the work address enrolled on the device.

In Settings > Browser extension, each feature is set to Forced for all users, Per-user choice (default) or Disabled.

SaaS visibility feature

• Browsing, on all sites, not only SaaS applications: the host of each page opened (https://host/), the type of navigation and the time. Never the parameters or the fragment. The path is kept only on chatgpt.com, chat.openai.com, copilotstudio.microsoft.com, zapier.com, agents.zapier.com, cursor.com, *.retool.com and *.app.n8n.cloud.
• Services contacted by pages: third-party hostnames, with the page's domain and a counter, except for a fixed list of technical hosts (audience measurement, CDN, advertising).
• Sign-ins to applications: the site, the methods offered and the one used, the context and the time. The address used is sent only if its domain is the domain of the enrolled address; otherwise, the "personal" or "external" label is sent instead. The password field is read only to know whether it is empty.
• Developer consoles (GitHub, Vercel, AWS, Google Cloud, Cloudflare): the type of secret, the count, the page address (the server removes its parameters) and the key reduced to its first 4 and last 4 characters, with no surrounding text.
• Extension inventory: ID, name, version, description, state, install type, permissions, allowed hosts and home page.

AI analysis feature

For each message sent to a covered AI tool (ChatGPT, Copilot, Gemini, Claude and other dedicated tools, and the AI panels of Gmail, Outlook, Google Docs, Microsoft 365, Notion, HubSpot and Zoom):

In the preview, keys and tokens, private keys and numbers with a strong format (card, SIN, IBAN, etc.) become a marker such as [REDACTED:credit_card]; your organization's high-severity patterns become [MASKED:…]; email addresses and phone numbers are also masked. A message of 200 characters or fewer is therefore kept in full, except for these items. If the server still finds a secret in the preview, it replaces the preview with a generic mention. The justification (2,000 characters at most) is not masked.

• Message cancelled or blocked: no text, only the detected types and the outcome.
• Files and images: analyzed on the device; only the name, the type and the result are sent.
• AI tool responses are never collected.

Always sent by a connected device

Whatever the features: the browser name and version, the extension version, a random device identifier, technical counters with no content, and requests for access to a tool or a site, with their justification.

Your organization's settings

• No excerpt: no preview is kept, but the 4,096-character buffer is still sent for re-analysis, then discarded.
• Excluded people and groups: no AI events from these people are reported; their safeguards stay active.
• Full text of flagged prompts: keeps, encrypted, the first 4,096 characters of flagged messages, masked on the device. Each read is logged.
• Password checks (Chromium only): the comparison happens on the device, and only the findings are sent. With the known breaches option, api.pwnedpasswords.com receives the first 5 characters of the SHA-1 hash.

Employee consent

• Manually installed device: the employee signs in with Microsoft, then confirms or declines each optional feature (Confirm my choices or Decline all), and can change their mind from the extension icon. A required feature is locked.
• Managed device (deployment key pushed): no choice is offered. Both features are active except the ones you disable: "per-user choice" therefore means "active". An information tab opens at enrollment; its read receipt is logged.

Once the device is enrolled, protection rules (block, warning, masking) apply on the device even if the AI analysis feature is not active; in that case, no data about that message goes to Sorvek.

Retention

• Previews, indicators and console secrets: 90 days by default, from 7 to 365 days depending on your Data retention setting.
• Full text of flagged prompts: 14 days by default, from 1 to 30.
• Browsing, contacted services and sign-ins: no guaranteed deletion period; this data may remain as long as the service is active.

All of it is deleted when the organization is erased. See also Data retained, retention periods and deletion and the extension privacy policy.