Security incident handling
For credential leaks, unauthorized access, malicious tools, or abnormal external actions, stop the impact first, preserve evidence, and reconcile results. Do not repeat harmful actions to demonstrate the problem.
Prioritize containment
| Observation | Immediate action | Verify |
|---|---|---|
| Repeated harmful scheduled actions | Pause the task and inspect execution records | Future triggers paused; reconcile active operations separately |
| Plugin or MCP exceeds scope | Disable the target connection in Extensions | Connection state and provider grants independently |
| A secret entered logs or a shared file | Revoke or rotate it at its service and stop sharing | Old credential invalidation and sent copies |
| Incorrect code changes | Preserve the diff, stop edits, and restore task-owned changes | Existing edits retained and behavior revalidated |
See automation and external actions. Do not repeat deletion, sending, or payment to investigate.
Submit Desktop feedback
Open application menu > Feedback, or use macOS Cmd+Shift+U / Windows Ctrl+Shift+U. Find version in About and diagnostics under Settings > Advanced > service details > Open log folder or Copy diagnostics. Review screenshots and attachments, then inspect submission status. Session /feedback drafts a template; it does not submit a report.
Include observed action, affected objects, version, OS, mode, executor, time and timezone, confirmed effects, containment results, and redacted errors. Feedback neither revokes credentials nor replaces provider account recovery. Contain the incident first and submit minimal evidence without customer directories or secrets.
Follow the response order
- Stop relevant tasks, schedules, or processes and disconnect the affected tool. Inspect queued and in-flight work.
- Revoke grants or rotate credentials at the owning service and confirm old credentials are invalid.
- Preserve time and timezone, task or node ID, tool source, target, actual diff, and redacted errors.
- Inspect files, remote messages, publications, charges, and account activity. Separate completed actions from unresolved outcomes.
- Remove the affected skill or tool, restore incident-owned changes, and validate within limited scope before resuming.
Prioritize by incident
Revoke leaked credentials first. Stop malicious extensions and inspect dependencies. Handle accidental sending or publication at the destination service. Stop software with an abnormal source and verify origin again. Git recovery handles files, not external effects.
Report safely
Use the relevant product or service support channel with minimal reproduction and redacted evidence. Include observed facts, affected objects, actions taken, and unresolved outcomes. Exclude full environments, tokens, private keys, and customer datasets. Correct the authorization, source-review, or data-handling failure before restarting automation.