We often hear the sentence “our backups are immutable” in meetings. The question that should follow is almost never asked: against whom? Before moving to encryption, a ransomware operator tries to reach the backup console, storage management or the cloud account, and each immutability technology closes only some of these paths. Below we compare S3 Object Lock, WORM and array snapshots, hardened Linux repositories and tape options by the privilege level they withstand and how they get breached in the field.
What immutability protects against
Immutable backup means a written backup file or object can’t be deleted or overwritten, and its period can’t be shortened, until the retention period ends. The real issue is the question “by whom?” The strength of a control is measured by the lowest privilege that can switch it off. Before starting the assessment, put these attacker profiles on the table one by one:
- Domain admin: Active Directory has been compromised; they have administrator rights on every domain-joined server.
- Backup console administrator: In the backup software’s interface, they have the right to delete jobs, shorten retention periods and remove repositories.
- Storage or operating system root: Root on the repository server, administrator on the array, or access to the server’s remote management card (iLO, iDRAC, IPMI).
- Cloud account owner: The highest privilege in the account holding the object storage has been taken over.
- Physical access or an insider: Someone who can physically take away the disks, cartridges or the device.
Let’s make one thing clear from the start: immutability doesn’t prevent reading. An attacker exfiltrating backups and using them for extortion is a separate risk on the enterprise cybersecurity side; it requires encryption and network controls. The only thing we’re discussing here is keeping the ability to recover intact.
S3 Object Lock: governance and compliance modes
Amazon S3 Object Lock applies WORM (write once, read many) protection to object versions. Versioning must be on for the bucket; the lock is tied not to the object itself but to a specific object version. A delete request sent without a version ID only adds a delete marker; the locked version stays in place. AWS defines two retention modes:
- Governance mode: Most users can’t delete a locked version. However, an identity with the
s3:BypassGovernanceRetentionpermission can delete the version or change its retention setting by adding the bypass header to the request. An attacker who takes over the account’s most privileged identity grants themselves this permission. - Compliance mode: During the retention period, nobody, including the root account, can delete the protected version or shorten the period; the period can only be extended.
In addition, there’s legal hold, which isn’t time-based: it prevents deletion until removed, and can be removed by a user with the relevant permission. The protection that matters in a ransomware scenario is compliance mode. Governance mode is a seat belt for testing and accidental deletion, not against an attacker.
Compliance mode has its blind spots too. An attacker can set very long retention periods on newly written objects and mount a cost attack. Closing the account, or its suspension over a payment problem, affects the availability of the data. With on-premises S3-compatible object storage appliances, things are different again: it isn’t enough for the API to offer compliance mode; someone with access to the appliance’s operating system, disks or management interface bypasses the API entirely. For these products, ask the vendor two questions in writing: how is compliance mode enforced against the appliance administrator, and how is the time source protected?
Which mode the backup software uses Object Lock in, and with how much extra time, varies by product. A common situation is that the lock is configured on the bucket but the written objects have no retention information at all. Check the bucket setting and individual objects:
aws s3api get-object-lock-configuration --bucket backup-bucket
aws s3api get-object-retention --bucket backup-bucket --key "example/object/path" --version-id VERSION_ID
WORM storage and storage array snapshots
There are two separate approaches in enterprise NAS and arrays. The first is file-level WORM: NetApp ONTAP SnapLock and Dell PowerScale SmartLock are in this class, and both offer two modes, “compliance” and “enterprise”. In enterprise mode, an authorized administrator can delete locked data under some conditions; compliance mode aims to protect against the array’s own administrator too. The second is locking array snapshots against deletion: features such as Pure Storage SafeMode and NetApp ONTAP’s tamperproof snapshot locking prevent a snapshot from being deleted with a single administrator account. How the protection is removed (a vendor support process, a second approval, the period expiring) varies by product and version; verify it in the vendor’s current documentation.
A snapshot gives fast recovery but doesn’t replace a backup. The snapshot sits on the same array; if the array fails, or the management plane is taken over and the array reconfigured, the snapshot goes too. Same box, same fate. When assessing, get answers to these questions: Whose approval is needed to remove the protection? If the array’s time source is changed, does the period end early? Does a factory reset or recreating the disk group get around this protection?
Hardened Linux repositories
In this model, backups are protected on a single-purpose Linux server with the file system’s immutability attribute. A common example is the Veeam Hardened Repository: the repository is added with single-use credentials that aren’t stored permanently on the backup server, Linux’s immutable attribute is set on backup files, and a local service running with root rights on the repository server sets and removes this attribute. Even if the backup console is compromised, the console can’t delete the files before the period expires.
The weak point is the repository server’s root account. Someone who reaches root removes the attribute or formats the disk. The paths to root are well known: SSH, the management card’s virtual console, physical access and the time source. Because the period is calculated by the clock, an attacker who can move the server’s clock forward tries to end the period early. The gap we often see is SSH left open after installation. The installation and hardening steps are described in order in the Veeam hardened repository installation guide.
Tape and offline media
Tape removed from the library and stored elsewhere is the only option that can’t be accessed over the network with any privilege; that’s a real air gap. LTO-WORM cartridges additionally prevent overwriting at the media level. But a cartridge sitting loaded in the library isn’t offline: an attacker who takes over the backup server can delete or overwrite it. With tape, protection comes not from the technology but from the discipline of physically removing the cartridge and storing it off-site.
The price of tape is recovery time and operational load. Restoring many servers from tape takes much longer than from a copy on disk. Cartridge rotation, inventory and restore testing rely on personal discipline. If the discipline slips, the tape stays in the library. You can find a comparison of tape and object storage in terms of long-term retention economics in the tape vs object storage guide on the systems side.
A comparison table by attacker privilege
The table shows, assuming each option is configured correctly, up to which privilege level it holds. “Partly” means the protection depends on an additional control (separate identity, separate network, second approval).
| Attacker privilege | S3 compliance mode (cloud) | WORM / protected snapshot | Hardened Linux repo | Offline tape |
|---|---|---|---|---|
| Domain admin | Protects | Protects (if the array is separate from the domain) | Protects (if the repo isn’t a domain member) | Protects |
| Backup console administrator | Protects | Protects | Protects | Protects |
| Repository/array administrator or root | Not applicable | Partly (depends on the mode) | Doesn’t protect | Protects |
| Cloud account root privilege | Deletion blocked; risk of account closure | Not applicable | Not applicable | Not applicable |
| Time source manipulation | Provider’s clock | Partly | At risk | Not applicable |
| Physical access | Provider’s responsibility | Doesn’t protect | Doesn’t protect | Depends on the storage location |
What the table says: a single immutable copy depends on a single privilege layer. For most organizations, a reasonable design has two copies. One immutable copy on-premises for fast recovery (a hardened repo or protected array), and a second copy in a different management plane (object storage in compliance mode or tape stored off-site). Separating these copies’ credentials from each other and from the domain is as decisive as the technology chosen; we cover this separately in the identity isolation in backup infrastructure article.
A short decision tree for choosing
- If your recovery time target is in the order of hours, the first copy should be on-premises and on disk: a hardened repo or protected snapshot.
- If you can’t separate the repository server and arrays from the domain and the shared management network, move the second copy to a different management plane: compliance-mode object storage in a separate account, or tape stored off-site.
- If the cloud account is managed with the IT team’s everyday identities, separate the account first; otherwise the cloud copy shares the root account risk.
- If regulations or contracts require long retention, consider tape or a low-cost object storage class for long-term (GFS) copies.
Retention period and capacity impact
The immutability period is a capacity load on top of the retention policy. Because data can’t be deleted before the period expires, two effects arise. First, incremental backups depend on the full backup they’re tied to; software keeps the entire protected chain for the duration, and some automatically extend the period by a few days to manage block reuse. Second, even if you shorten the retention policy, space isn’t freed immediately. When the disk fills up, this second effect comes as a surprise.
An example calculation: Consider 14 days of immutability in an environment with 10 TB of source data, where daily incremental backups take up 0.5 TB on average and the weekly synthetic full is taken with fast clone. If fast clone isn’t supported, each weekly full adds about 10 TB, and within the two-week window at least two fulls and 14 incrementals (about 27 TB) stay locked. With fast clone (XFS reflink or ReFS block clone), synthetic fulls share unchanged blocks and physical consumption drops noticeably. The choice of file system is part of the capacity decision.
Object storage brings extra cost items: old versions kept because of versioning, minimum storage durations and early deletion fees in some storage classes, and the number of API requests. In compliance mode, a policy accidentally given a very long period can’t be undone until the period expires. There’s no going back. Try the policy in a short-lived test bucket before putting it live.
Pre-implementation checklist
- For each immutable copy, there’s a written answer to “who can remove this protection, with which identity?”
- The credentials for the repository, array and cloud account are independent of the domain and the backup console.
- The time source (NTP) is protected and clock drift is monitored.
- The immutability period was set taking into account how long it may take to notice an attack; a period of a few days isn’t enough in most scenarios.
- The protection was actually tested: deletion attempts from the console and from the operating system were refused.
- At least one copy is in a different management plane or offline.
An immutable copy alone isn’t a recovery plan; choosing a clean recovery point and the restore order must also be prepared in advance. For the general framework, see the ransomware recovery readiness article. Assessing your current backup architecture against this table together is the first step of our backup security and ransomware readiness work. The primary source for the details of Object Lock behavior is the AWS S3 Object Lock documentation. To turn the chosen method into a workable recovery architecture, see our enterprise backup and recovery infrastructure service.
Frequently asked questions
Is governance mode in S3 Object Lock enough against ransomware?
No. In governance mode, an identity with bypass permission can delete locked versions; an attacker who takes over the account’s most privileged identity can also grant themselves this permission. Use compliance mode in a ransomware scenario, and leave governance mode for testing and protection against accidental deletion.
Should I choose a hardened repository or S3 compliance mode?
The two close different risks. A hardened repository gives fast on-premises recovery but depends on the repository server’s root account and management card. Object storage in compliance mode sits in a separate management plane, while restore time depends on your connection. If possible, use both together, with different identities.
How many days should the immutability period be?
There’s no one-size-fits-all value. The period should cover the time that may pass until the attack is noticed and the need to search backwards for a clean recovery point. Periods of a few days are too short in most scenarios; verify the capacity impact with a sample calculation and then decide.
Do tape cartridges sitting in the library count as offline?
No. Cartridges loaded in the library are accessed through the backup server; an attacker can delete or overwrite them. For offline protection, the cartridge needs to be removed from the library and stored in a separate location.