MAQPNA

Learn

How to stop an AI agent deleting a production database.

Telling the agent not to is not a control. Taking away its ability to do it is.

Updated 31 October 2026.

FAQ

Common questions

Can a system prompt stop an AI agent deleting data?

Not reliably. In the Replit and PocketOS incidents the agents had instructions not to make such changes and acted anyway. Treat prompt rules as guidance, and enforce limits in credentials and in a policy check the agent cannot bypass.

Should AI agents ever have production database access?

Read access is often fine. Write and delete access should be scoped to the task, short-lived, and gated by approval for anything destructive. Many teams give agents no direct production write access at all.

Is a soft delete or a delay on deletes enough?

It helps recovery, as Railway's 48-hour window shows, but it does not stop the action. Combine it with least privilege, approvals and backups the agent cannot reach.

Does MAQPNA replace database permissions and backups?

No. MAQPNA governs the tool calls an agent makes through its gateway. Database roles, environment separation and backup design remain your responsibility, and they work best together.

Put a gate in front of every destructive call.

Run your agent under a local MAQPNA, write one deny rule and watch the gateway refuse the call.

More install options
curl -fsSL https://maqpna.com/install.sh | sh