Browser Capture
AI chats in the browser are captured by the Waxell browser extension. This page covers the controls for it, all on the Guard's Browser Capture tab. Like every Guard setting they can be set for the whole workspace and overridden for a team, an app type or a single device (see The Guard Cascade).
Once the Waxell agent and extension are installed, browser AI conversations on the built-in sites are captured and uploaded, with secrets redacted on the device first (see Privacy & DLP). To keep captures off your workspace, turn Browser capture off on the Global layer before you deploy.
Turn upload on or off
Browser capture decides whether captures leave the device.
- On (default) — captures upload and appear in Conversations and Device Activity.
- Off — the extension keeps recording, but the agent holds captures on the device instead of uploading them. Turning it back on delivers what was held. The Health page shows held for the device.
Use it for a privacy-sensitive team, a pilot that should start metadata-only, or a paused rollout.
Choose which sites are covered
The extension covers ChatGPT, Claude, Gemini, Microsoft Copilot and Perplexity
out of the box. Add any other AI site's hostname to Browser extra hosts
(for example chat.mistral.ai). On macOS, re-download and re-push the browser
profile from the Setup page after changing this list.
How a site's send, reply and account are recognised is defined by its site adapter (Endpoints → Site Adapters). Override one there if a site changes its page and captures stop arriving.
Enforce browser DLP
Browser DLP mode checks each prompt before it is sent:
| Mode | Effect |
|---|---|
off (default) | Nothing is stopped. Captures are still redacted before upload. |
warn | The user is told what was detected and may continue. |
block | The send is refused with a message saying why. |
Browser DLP types chooses what counts — by default API keys, bearer tokens, SSNs, card numbers, cloud and SaaS credentials, private keys, JWTs, connection strings and labelled secrets. Email addresses are not included by default. Browser DLP hosts limits enforcement to particular sites (empty means every covered site). Browser DLP patterns adds your own detectors or replaces built-in ones — for example an internal token format.
If the extension has not received policy yet, it lets prompts through rather than blocking everything.
Scan file uploads
File upload mode checks files attached to AI chats on the device:
| Mode | Effect |
|---|---|
off (default) | Uploads are not scanned. |
scan | The upload proceeds and the file is scanned on the device; findings are recorded. |
strict | The upload is held until the scan passes, and refused if the file contains a secret or cannot be read. |
Files larger than File upload max bytes (25 MB by default) are not opened:
scan lets them through and flags them, strict refuses them.
Native apps and TLS interception
Earlier versions of Waxell could capture native desktop apps by intercepting their encrypted traffic. That capability is disabled in the current Waxell agents, and the Browser Capture tab shows it collapsed and labelled Disabled in agents. Turning it on has no effect on devices and does not change browser capture. Native coding agents are governed through hooks instead, and the ChatGPT desktop app through its own capture setting.
Check it worked
- Open an AI site on a managed device and send a message.
- On Health, the device's Capture column should read on.
- The conversation appears in Captures and in Device Activity for that device within a few seconds, with any secrets shown as redacted.
If the device reads no ext, see Confirming the extension is working.