Privacy, data handling and when to stop an agent

Privacy by design

An agent can only expose what it can see, so privacy mostly means showing it less.

• Minimum data. Give the agent the narrowest key, one mailbox rather than all, and read-only access unless writing is needed.
• Minimum retention. Memory holds notes between runs. Save only what is useful, never secrets or personal details you do not need, and use Forget on the Memory panel to remove notes.
• Minimum destinations. Limit web calls to Allowed web domains and list only the connected servers it needs.
• People's data. If an agent handles information about other people, such as customers or colleagues, apply the same care you would to the original data, and make sure you are allowed to process it that way.

What is kept

PGB keeps a copy of what you type into AI features, including an agent's instructions and the input of each run, for a limited period, to investigate abuse such as phishing and malware. Secrets are masked before saving, your files, mail and other data that an agent reads to answer you are not part of that copy, and the copy is not used for advertising or to train models. The Safety card on Settings and the guide "How agents stay safe and what is kept" explain it in full. For you it means: do not put secrets or sensitive personal data in instructions or in run inputs.

Errors you can show others

If something unexpected fails, the run shows a short message with a reference starting with ERR-. That reference lets support find the details without exposing internals. Share the reference, not screenshots of private data.

When to stop an agent

Stopping is a skill. Stop when:

• It did something you did not intend, even once, and you still do not know why.
• It is spending faster than expected.
• A secret, key or webhook address may have leaked.
• The job no longer exists, or its inputs changed in a way you have not checked.
• Nobody is watching its results any more.

The stop ladder

From gentlest to firmest:

1. Cancel the run with Cancel.
2. Pause the agent with Pause (then save) and Save changes. Schedules and webhooks do nothing while it is paused.
3. Cut access. Remove the key under PGB Business access, Remove a secret, Rotate secret on the webhook, and revoke the key in PGB Business.
4. Investigate. Open Runs and read the traces. Check Reports for when the change began.
5. Restore or delete. Use Versions to go back, or Delete agent to remove it with its runs and memory after confirmation.

For a suspected leak, do step 3 first, then step 2.

Try it yourself

Write a one-page "stop plan" for an agent you built: the three signs that would make you pause it, and the exact buttons you would press, in order.

Recap

• Privacy is mostly about showing agents less: narrow keys, small memory, few destinations.
• Keep secrets and sensitive data out of instructions and inputs.
• Stop early: cancel, pause, cut access, investigate, then restore or delete.