In most environments, an administrator with normal backup platform access can erase the backups it produced in under an hour, usually without tripping an alarm. That makes backups an obvious target for an insider threat: a rogue administrator, a careless one, or an attacker using an administrator's login. Whoever can set retention can also change it.
Immutability is often described as if it settled the question. It closes real paths, but not all of them, and the ones it leaves open are those an insider is best placed to use. Separating what is blocked from what is merely inconvenient is the only way to know which risks have been accepted rather than removed.
Three routes through the backup console
- Deletion. An administrator selects restore points, or a whole chain, and removes them. Well designed platforms log this and may ask for confirmation, but the operation exists because legitimate cleanup exists, and the log is written to a system the same administrator can usually reach.
- Retention. Shortening a retention period deletes nothing immediately. The next retention pass removes everything now considered expired, through the same routine process that clears old data every night. The log shows a policy edit, not a mass removal.
- Repository removal. Removing the repository from the console, or reformatting the volume beneath it, discards the data without any per object deletion. Where the repository is a general purpose file system or block device, this is a storage operation rather than a backup operation, and the software may only notice afterward.
Routes that bypass the backup software entirely
If backups land on object storage, the S3 credential used by the backup jobs also works directly with a command line client. What that credential can do is defined by storage policy, not by the backup software, and a key granted full access can delete objects the console would have refused to remove.
One step further down, an identity able to edit bucket policy, or to change versioning and lock configuration, can alter the rules protecting objects before deleting them. The backup platform's audit trail will contain nothing about it. Below that sits the infrastructure: a hypervisor administrator can delete the virtual machine that is the repository, and someone with access to the array or appliance management interface can destroy the underlying volumes. Physical access needs no credentials at all. These layers are often run by the same small team, which is the practical case for separating backup administration from general infrastructure administration and for measuring each account's administrative blast radius.
| Deletion path | What it removes | What closes or narrows it |
|---|---|---|
| Console deletion of restore points | Selected backups or a whole chain | Object Lock on the storage side, since the storage refuses the delete regardless of console intent |
| Retention shortened in the backup software | Everything newly considered expired, at the next pass | Retention enforced on the object rather than by the job, so the lock outlives the policy edit |
| Repository removal or volume reformat | The full contents of that repository | A second copy on separate storage under separate credentials |
| Direct use of the storage credential | Objects the policy allows that key to delete | A narrowly scoped key without delete rights and without the right to lift a lock |
| Bucket policy or lock configuration change | The protection over everything in the bucket | Compliance mode locks, plus a separate identity for bucket administration |
| Hypervisor or hardware layer destruction | The platform the repository runs on | Physical separation of one copy, and different administrators at that layer |
What Object Lock blocks against an insider, and what it leaves open
S3 Object Lock closes the first two rows cleanly and the fourth partly. A locked object version cannot be deleted before its retention date, so console deletion fails at the storage layer, and a shortened retention policy in the backup software does not shorten the lock already written to the object. A misused storage key cannot delete locked objects either, provided it lacks the right to change lock settings.
Whether a lock can be lifted early depends on its mode. Mode and retention period apply to each object version, usually inherited from a bucket default. A governance mode lock can be overridden by an identity holding the s3:BypassGovernanceRetention permission. A compliance mode lock cannot be removed or shortened by anyone, the root user included, until it expires, according to Amazon's S3 Object Lock documentation. That difference is the difference between an administrator who can reverse a decision and one who cannot, covered further in governance mode compared with compliance mode.
Two things stay outside the lock. The credential a backup job writes with and the identity that administers the bucket are separate objects with separate policies on any S3 platform; whether one person holds both is a local decision, and it decides whether the bucket policy row is open or closed. And a lock is enforced by software running on hardware, so an administrator of that hardware is outside its scope. That is the boundary where immutability and physical separation stop being interchangeable, as an air gap compared with immutability explains.
The day-to-day cost of each control
Every control in the right hand column has an operating price, and ignoring it is how controls get removed six months later. Compliance mode means capacity cannot be reclaimed early: data written under a three year lock occupies storage for three years even if the workload it protected was retired in month four.
Splitting administration costs coordination. If the person who runs backups cannot administer the bucket, a genuine mistake, such as a job writing a very large volume under a long lock, needs a second person to resolve and sometimes cannot be resolved at all. Teams of three or four feel this immediately.
A separate copy on separate infrastructure costs money and attention. It is also the only control that survives every row of the table, which is why the 3-2-1-1-0 structure keeps one copy outside the main environment.
Mapping the paths in one environment
For each row of the table, write down whether the path is open in the current environment and who could use it. The exercise takes an afternoon and usually produces two or three names. That list is the real answer to the question in the title, and it is more useful than any general statement about immutability.
Close paths by cost. Narrowing the storage key is cheap and rarely regretted. Splitting bucket administration is worth it where headcount allows. Closing the hardware path needs a copy somewhere else and is a budget conversation. Review the list whenever someone changes role or leaves, and date each review so staleness is visible.
