Privacy & DLP
Endpoints records AI use so you can govern it. This page sets out exactly what that record contains, what never leaves the machine, and the controls you have.
The commitments
- Only AI use is captured. The browser extension runs on AI sites — the built-in ones and any you add — and nothing else. Banking, mail, health and every other site are never captured.
- No network decryption. Waxell reads AI conversations inside the browser page and receives coding-agent activity from the tools' own hooks. It does not intercept or decrypt network traffic. (That capability existed in earlier versions and is disabled in the current agents.)
- Secrets are redacted on the device. Before a capture is uploaded, the agent removes credentials and sensitive identifiers. Waxell never receives the original values.
- You control whether captures leave the device, for the whole workspace, a team, an app type or a single device.
- Every capture is attributed — device, user, app, AI site and account type — so the record is evidence, not a pile of text.
What is captured
| Source | What is recorded | Default |
|---|---|---|
| Browser AI chats | Prompt, reply, model, site, account type (corporate / personal / unknown) | On |
| Coding agents (hooks) | Commands, file paths, tool calls, prompts, and each allow / ask / deny decision | On wherever hooks are installed |
| ChatGPT desktop chat | Prompt and reply, read after the fact | Off |
| Network metadata | Which app reached which AI host, when, as which user — no content | On |
What is redacted before upload
API keys and bearer tokens; AWS, GitHub, Slack, Google, Stripe and Waxell
credentials; private keys; JWTs; database connection strings; labelled secrets
(password=…, token: …); payment card numbers (checksum-validated); US Social
Security numbers; and email addresses. Each is replaced with a marker naming its
type, and the capture is labelled redacted.
Add detectors for your own formats — an internal token, an employee ID — with Browser DLP patterns on the Browser Capture tab. They apply to redaction as well as enforcement.
Holding capture on the device
Turn Browser capture off on any Guard layer and captures from those devices stay on the device instead of uploading. The device keeps a local trail; turning upload back on delivers it. Typical uses: a works council or legal requirement for one region, a privacy-sensitive team, or a pilot that should start with metadata only.
For one device in a hurry, use Capture break-glass on the device's page; it overrides every layer for that device.
Redaction is not blocking
Redaction protects Waxell's stored copy. It does not stop the user sending a secret to the AI service. To stop that, turn on enforcement:
- Browser DLP mode (
warnorblock) for browser chats — see Browser Capture. - Agent DLP mode for coding agents — see Coding Agents (Hooks).
Retention and access
Captures are stored in your workspace and visible to your admins under Endpoints → Captures and Device Activity. Who can see them follows your workspace roles.
How long captures are kept is set under Settings → Data & Privacy → Retention (Endpoint & browser captures), from 30 days up to 7 years.
Related
- Browser Extension — how capture works in the browser.
- The Guard Cascade — scoping any of these controls.