Security and data

Nothing in the Workbench collects from your machine on its own. This page documents what moves where, so you don't have to guess.

What never happens

• No passive collection. Tools run only when you (through your AI client) ask for them.
• The UE connector makes zero outbound network calls and holds no credentials — it only registers local tools and writes settings inside your project.
• No API keys or long-lived tokens are stored in your client config. Authentication is OAuth handled by your client, with short-lived access tokens it refreshes on its own.

Where your files go

When a tool needs a file — a log for analysis, an inventory for an audit:

• Your client uploads it directly to object storage through a presigned link scoped to that one session. File bytes never flow through the chat or the model context.
• The link expires after 15 minutes; the session it belongs to after 24 hours.
• Analysis runs server-side against the stored copy. Report download links expire after 24 hours.

What gets stored

• Ingested log events from analyses you run — parsed, with emails and tokens redacted, tagged to your session and project. These power the temporal research tools.
• Session and usage records — what ran, when, and what it billed. These power the portal's Usage and Billing pages.

Controls that keep you in charge

• Credits gate before work — a billed tool can't start unless the cost is covered, so nothing runs "by accident". See How credits work.
• Human approval for generated scripts — make_tool results are dry-run transforms and become real only when you approve them; destructive operations are commented out. See make_tool.
• Session ownership — a session is usable only by the account (and team) that created it; foreign sessions are rejected.

Next steps

• How it works
• Fix a problem