Backup recovery points cannot be deleted by AI
Author
Audrey
NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.
Date Published
Backups are the last line. When something goes wrong in production, whether it is a bad deploy, a deleted database, or an intruder, the recovery point is what gets the team back. People who attack systems know this, which is why deleting backups is a standard early move.
Now put an AI agent in the operations toolkit. It can list resources, clean up what looks unused, and tidy storage. Somewhere in that toolkit sits a tool that deletes recovery points, because the cloud and backup platforms offer one. The agent has it for the same reason it has everything else: someone connected the platform.
The problem in IT's terms
The infrastructure lead has two scenarios in mind, neither of which requires malice.
- A helpful agent asked to "clean up old snapshots to save money" decides that recovery points are old snapshots.
- A session is steered by text it read, in a ticket or a log line, to remove them.
Either way the recovery points go, and the team finds out when they need them. Backup retention policies on the platform help, but an agent with a delete tool is a separate path around them. The lead wants that path closed at the tool, for every agent.
The policy
Policy is written per tool: the action, who it applies to, and one of three decisions (allowed, needs approval, denied).
Action | Who | Decision |
|---|---|---|
List backups and recovery points | Operations and platform agents | Allowed |
Run a backup job | Operations agents | Allowed |
Delete a recovery point, in the backup service or the backup appliance | Every agent | Denied |
The delete tool is denied for all agents, in every app and for every user. There is nothing for a clever prompt to unlock, because the answer is the same each time. The agent can still do the useful work around backups: listing them, reporting on age and size, and starting jobs.
Who handles it instead
If a recovery point really does need to be removed, a person does it through the platform's own console or pipeline, the way the team already handles retention exceptions. That keeps removal a deliberate human action with a human's name on it.
A team that prefers a different shape can set the tool to needs approval instead. The request then goes to a named approver or role, who answers in Slack or the web inbox. For recovery points many teams will keep deny, because there is rarely a time-sensitive reason for an agent to delete one.
What the Record of Decisions shows
The Record of Decisions keeps one entry per decision: who asked, which agent, which tool, the decision, who approved, and when. For this scenario the entries worth reading are the refusals.
- Which agent called the delete tool, for which user, and when.
- Whether it happened once, as a cleanup gone wrong, or repeatedly, as a sign of something steering the session.
- Whether the attempts line up with a ticket, a log entry, or an integration that the team should look at.
A refusal that appears in the record is a small, early piece of information. It tells the team that an agent reached for a tool it should not have, while the recovery points are still there.
Use it as the working record of policy decisions alongside the backup platform's own logs.
What this does not do. It does not protect against everything that can happen to a backup. A person with console access can still delete one, and the platform's own retention and immutability settings are still worth using. This closes one path, the agent path, and records the attempts.
Related scenarios
- Restores ask and leave a record: restoring a database pauses for approval and is recorded.
- Plan yes, apply no: agents may plan changes in production, while applying them is denied.
- No public S3 buckets: another storage action denied to every AI.
- Backup testing evidence for examiners: the record of restore tests the backup agent ran.
Get started and set the same line on your backup tools.

