Agents work from your briefs, your documents and live web data, so their outputs sometimes contain things that shouldn't travel: a customer's email address, a card number, an API key pasted into a config file the agent read. Forge scans every output for these. What happens next is your organisation's choice.
Step 1: Open your organisation settings
Go to Settings → Organisation and find Sensitive data in outputs. You need to be an organisation admin: the setting governs everyone's runs.
Step 2: Choose what should happen
Flag only — detect it and show it on the run, changing nothing. This is the default and what Forge has always done.
Redact — replace matches with placeholders like [REDACTED:email] before the output is stored. The sensitive text never lands in your data at all.
Block credentials — redact personal data as above, and withhold entirely any output containing a live key or token.
Step 3: Check a run
When something is redacted or withheld, the run says so and names what was found. If you see the same category appearing repeatedly, the fix is usually upstream — a document in your knowledge base, or a brief that keeps including live data.