Keep assistant access safe
Limit what an assistant can do, review its activity and cut it off.
Decide how much access
| If you want | Do this |
|---|---|
| Assistants that only look | Use a sign-in with a view only role, or a token with Read areas |
| Assistants that make changes | Give edit access only to the areas needed |
| No sign-in connections | Turn the organisation switch off in Integrations > AI assistants |
Sign-in connections follow the approving user's role. Token connections follow the token's areas. Finance tools are read only, and billing changes need a role with full finance access. See Permissions.
Review activity
The activity feed in Integrations > AI assistants lists tool calls. Filter by reads, writes or connection events, by user, connection or token, by date, or by failed calls only. Each entry shows what was called, who called it, and whether it came through sign-in or a token.
Check writes regularly while you are getting started.
Cut off access
- Revoke one connection or all connected apps. A revoked app is cut off on its next call.
- Turn off the organisation switch to cut off every sign-in connection straight away.
- Disable or rotate a token to cut off a token connection.
- A sign-in connection that isn't used expires after 30 days.
Good habits
- Start read only and add write access when you trust the setup.
- Use a dedicated token for each assistant or tool.
- Ask the assistant to describe a change before it makes it, and use
dry_runwhere tools offer it. - Remember that what an assistant reads goes to that assistant's provider. Check that your provider and privacy terms suit client information.
Was this page helpful?