Stopping an AI agent from deleting production.
Agents now touch real systems: code, servers, databases, deployments. There have been public reports of agents deleting production data while holding broad access, with nothing checking the specific action. These five rules are the checks a person would expect.
Five rules for agents that touch production.
These are standard change-control practice, applied to the agent. They are a starter set for production change, built on Torvant’s general rule engine, not the whole platform. Every decision is kept, refusals included.
Production changes need a named approver
An agent changes a live system because it held broad access and nothing asked a person first.
- What you get
- Set the number of approvers once. Every production change waits for them, destructive changes need 2 approvers, and each decision is on the record.
Freeze windows the agent cannot lift
An agent keeps changing things during a code freeze because nobody told it, or it decided the freeze did not apply.
- What you get
- Set a weekly freeze window once, at a fixed UTC offset such as +05:30. During it the checkpoint refuses changes, and the agent has no way to switch the freeze off. Freeze windows are a prototype.
Only the credentials the grant names
An agent finds an unrelated access token and uses it for something nobody gave it.
- What you get
- The agent gets keys only for the providers its grant names. Anything else is refused, and the refusal is recorded.
Caps on how much one agent can touch
One run deletes thousands of files or rows before anyone notices.
- What you get
- Write the limit into the grant once. The call that crosses it is refused, whether it is rows touched, files deleted or calls per minute.
No force-push without a person approving that exact push
An agent rewrites shared history on a branch the team depends on.
- What you get
- Keep branch protection on. Torvant adds the record of who approved which push.
Early product. Nothing is in production use yet.
The rules hold on structured calls that name the environment, the credential or the count as fields. A rule that reads text inside a free-form shell command or SQL statement is a speed-bump, not proof, because an agent can phrase around it. Credential and approval rules also need the tool to sit behind the checkpoint. What is built and what is not is on Security and status. See also ten ways agents go wrong and what an authority layer is.
See what these rules would have held.
Book a demo and we will replay a week of your agent logs to show what these rules would have allowed, held and refused. Freeze windows are not part of the replay yet. Nothing is blocked.