Blog/Enterprise security

Backup and disaster recovery: 3-2-1, immutability and testing

The 3-2-1 rule, why ransomware forced immutable backups, how to set RTO and RPO honestly, and why an untested restore is not a backup.

CyberWatch AI21 November 2026 · 11 min read
External hard drives and cables on a desk next to a laptop

The backup job had been green for four years. Nightly, to a network share on a server in the same rack, with a copy syncing to cloud storage using credentials stored on the backup server itself. When ransomware ran, it encrypted the production file server, then the network share, then propagated through the sync to the cloud copy. The restore the company had been paying for, and monitoring, and ticking off in its annual review, was gone in the same forty minutes as everything else.

Nothing was misconfigured by the standards of 2015. The design assumed the enemy was a failed disk, a fire or a deleted folder. It had no concept of an adversary with administrative credentials deliberately looking for the backups.

That shift is the single most important thing to understand about backup design now.

The 3-2-1 rule, and why it is only a start

The old rule is still a reasonable baseline:

  • Three copies of your data, counting the production copy.
  • Two different media or storage types, so one technology's failure mode does not take everything.
  • One copy off site, so a fire, flood or theft at one location is survivable.

It is a good rule for accidents. It says nothing about an attacker, which is why practitioners now commonly extend it to 3-2-1-1-0: three copies, two media types, one off site, one offline or immutable, and zero errors in a verified restore test.

Those last two additions carry most of the modern value. Everything else on this page is detail about how to achieve them.

What ransomware changed

Modern ransomware operators spend days or weeks inside a network before deploying anything. During that period they look for the backup infrastructure, and MITRE ATT&CK records the resulting behavior as a deliberate technique: inhibiting system recovery. The specific actions are consistent:

  • Deleting shadow copies and local snapshots on every host they reach.
  • Logging into the backup console with harvested administrative credentials and deleting retention points, or reducing retention to a day so the backups expire on their own.
  • Encrypting the backup repository directly, because it is reachable on the network and frequently holds the largest, most valuable data set.
  • Using stored cloud credentials found on the backup server to purge or encrypt the off-site copy.
  • Waiting out the retention window where they are patient enough, so that by the time encryption runs, every available restore point already contains their foothold.
  • Stealing the data first, so that even a perfect restore does not remove the leverage. A restore fixes availability; it does not unpublish anything.

Three design consequences follow.

First, the backup system is a separate security domain. If your domain administrator can delete your backups, then an attacker with your domain administrator account can delete your backups. Separate credentials, separate identity provider where feasible, phishing-resistant multi-factor authentication, and no shared accounts with production.

Second, immutability is not optional. Policy that says backups should not be deleted is not a control. Storage that refuses to delete them is.

Third, retention has to be long enough to reach back before the intrusion. If your oldest restore point is fourteen days old and the attacker was inside for six weeks, you have no clean copy. Retention is a security parameter, not just a cost parameter.

Our guides to malware and ransomware and how ransomware gets in cover the attack side, and CISA's #StopRansomware Guide is the practical reference for the response.

Immutability and air gaps, in plain terms

The aim is a copy that cannot be destroyed by someone with your credentials. There are several ways to get there, and they differ in cost and in how convincing they are.

  • Object storage with a retention lock. Cloud object storage can hold data under a lock that prevents deletion or modification until a date passes, enforced by the storage service and not by your backup software. In compliance mode, nobody including the account owner can shorten it, which is the point. Check carefully which mode you have enabled, because a governance-mode lock that a privileged user can override offers much weaker protection.
  • Backup appliances with hardened immutable repositories, where the storage layer refuses delete operations for the retention window and the appliance operating system is not domain-joined.
  • Service-provider managed backup where deletion requires a separate out-of-band authorization, so your own compromised credentials are insufficient.
  • Genuine offline media. Tape, or rotated external drives that spend most of their life physically disconnected. Unfashionable, cheap, and still the hardest thing for a remote attacker to touch. The weakness is human: rotation gets skipped, drives stay plugged in, and nobody verifies the contents.
  • Logical air gap, meaning a copy on infrastructure reachable only through a separate authentication path, in a separate cloud tenancy or account, with no credentials for it stored on the production network.

Whichever route you take, test the thing that actually matters: try to delete a protected backup using your highest-privilege account. If it succeeds, your immutability is a setting someone forgot to apply. This is a five-minute test that very few organizations have run.

Sync is not backup

Worth stating plainly because it is the most common misunderstanding. A folder that syncs to cloud storage propagates every change, including encryption and deletion, to every other copy. Versioning and recycle bins in cloud platforms do help, and you should turn retention up as far as your licensing allows, but they are usually limited in duration and an administrator can purge them. Treat built-in cloud retention as one useful layer, not as the backup. Our cloud security guide covers what the platform does and does not do for you.

RTO and RPO: the two numbers that drive the design

These get quoted and rarely decided properly.

Recovery time objective is how long a system can be unavailable before the consequences become unacceptable. Recovery point objective is how much recent data you can afford to lose, expressed as a time window.

Both are business decisions, not technical ones, and they should be set per system rather than for the organization as a whole. The order of work matters:

  1. List your systems and what each one supports operationally.
  2. Ask the business owner two questions. How long can you work without this, and what would it cost you to re-enter a day of lost work? Expect "zero" as the first answer and then present the cost of zero.
  3. Group systems into tiers. For example: under four hours, under one day, under three days, best effort. Three or four tiers is plenty.
  4. Design backup frequency and recovery method per tier. A one-hour recovery point objective means snapshots or continuous replication, not a nightly job. A one-week objective means a nightly job is fine.
  5. Write down the actual measured recovery time, from your last test, next to the objective. The gap between the two is the finding.

Two things organizations consistently miss. Recovery time includes everything, not just the data copy: procuring or rebuilding hardware, restoring the operating system and application, reconfiguring, reconnecting dependencies and validating. And recovering one system is not the same as recovering fifty at once, which is what a ransomware event requires. If you have restored one server in three hours, a hundred servers will not take three hundred hours but it will not take three either, and the dependency order will matter enormously.

NIST's contingency planning guide is the standard reference for structuring this, including the business impact analysis that produces the tiers.

What to back up, beyond the obvious

File data is the part everyone remembers. The things that make a rebuild slow are usually the ones nobody listed.

  • Databases, using the database's own backup mechanism rather than copying files from under a running service.
  • Full system or virtual machine images for anything you cannot rebuild quickly from scratch.
  • Identity and directory configuration. Users, groups, permissions and policies. Rebuilding an identity system from memory while under pressure is among the worst experiences available in this profession.
  • Email and collaboration data, including chat history where it has business or legal value. Check what your provider's native retention actually covers, and for how long.
  • Network and security device configurations. Firewall rules, switch configurations, VPN settings. Exported and stored outside the devices.
  • Certificates, encryption keys and secrets, stored so they are recoverable but not sitting in the backup in plain form.
  • Source code and build pipelines, which includes the pipeline definitions, not just the repository.
  • Cloud platform configuration, ideally held as code in version control so it can be redeployed rather than reconstructed by clicking.
  • The recovery documentation itself. Printed, or stored somewhere that does not require the systems you are recovering. A recovery runbook on the encrypted file server is a lesson, not a plan.

Testing restores: the part that is always skipped

A backup you have never restored from is a hypothesis. Most organizations discover this during an incident, which is the most expensive possible moment.

A testing rhythm that is actually sustainable

  • Monthly: restore a handful of individual files chosen at random, from different dates in the retention window, and open them. This catches silent corruption, failed jobs reporting success and exclusions nobody noticed.
  • Quarterly: restore one critical system end to end into an isolated environment and verify the application works, not merely that the data copied. Record the elapsed time.
  • Annually: run a full scenario. Multiple systems, in dependency order, with the people who would really do it, assuming the production environment is unavailable and the documentation is the only reference.
  • After any significant change to the environment or the backup configuration.

What a good test deliberately includes

  • Restoring from the immutable or offline copy, not the convenient local one. That is the copy you will actually be using.
  • Restoring from an older point, not just last night's, because in a ransomware event you will need a pre-intrusion copy.
  • Verifying integrity and completeness, with the application owner confirming the data looks right, not just that files exist.
  • Working without the usual person. If only one engineer can run a restore, the plan has a single point of failure with a passport.
  • Working without network access to production, including without the password manager if that lives in the affected environment. Keep break-glass credentials available offline.
  • Recording the time each stage took, which is the output that matters most.

Then record the result somewhere durable, including what failed. A test with no findings usually means the test was too easy.

Received a message claiming your files are encrypted? Paste it into CyberWatch AI Scan for a free check before responding to anything in it.

Recovering for real: order of operations

The most common and most costly mistake in a ransomware recovery is restoring into an environment the attacker still controls, which produces a second encryption event days later. Sequence matters.

  1. Contain first. Isolate affected systems, disable compromised accounts, revoke sessions and tokens. Our incident response guide covers the full sequence.
  2. Protect what remains. Verify the backup infrastructure is isolated from the compromised environment and confirm which restore points exist and are intact, before touching anything.
  3. Establish the intrusion start date as closely as you can, so you know which restore points are safe to trust.
  4. Build a clean environment. New or rebuilt infrastructure, on a segmented network, with new credentials. Do not reuse the domain the attacker had administrative control of without a deliberate decision about how to make it trustworthy again.
  5. Fix the entry point before restoring, and reset credentials broadly, including service accounts.
  6. Restore in dependency and business-priority order. Identity, then core infrastructure, then the systems the business named as tier one. Validate each before moving on.
  7. Scan restored data and watch for reinfection. Keep heightened monitoring running for weeks.
  8. Handle the data theft separately. Stolen data is a breach with notification implications regardless of how well the restore went. See data breach notification.

NIST's guidance on recovering from ransomware and other destructive events, and its ransomware risk management profile, both set out this separation between restoring data and restoring trust in the environment.

Where the money should go

If the budget is limited, this is the order that buys the most resilience:

  1. One immutable or genuinely offline copy with credentials that do not exist in production. Nothing else matters as much.
  2. Retention long enough to reach back further than a plausible dwell time. Months, not days.
  3. A tested restore, with a measured time written down, for your most important system.
  4. Separate authentication on the backup system, with phishing-resistant multi-factor authentication. See administrator account security.
  5. Monitoring that alerts on backup failures and on changes to retention or deletion of restore points. An attacker reducing retention should trigger an alarm, and in most environments it triggers nothing.
  6. Printed recovery documentation with the contact numbers you would need, including your insurer and any incident response retainer. See cyber insurance.
  7. Coverage of configuration and identity, not only files.

A small-organization version

For a business with a handful of servers or none, running on cloud services:

  • Turn cloud platform retention and versioning up to the maximum your licensing allows, and know exactly how long it is.
  • Add a third-party backup of your cloud mail and files into separate storage with its own credentials and immutability. This is the gap most small organizations have.
  • Keep one encrypted external drive rotated weekly and stored off site, disconnected.
  • Restore five files from it every month and open them.
  • Write one page: what is backed up, where, who can restore it, how long it takes and the account recovery details. Print it.
  • Separate the backup account from your day-to-day administrator account.

That list is affordable and covers the majority of realistic scenarios. Our business cybersecurity guide and security on a small budget articles cover how it fits with the rest.

The short version

Three copies, two media types, one off site, one immutable or offline, and a restore test with zero errors. The backup system needs its own credentials and its own authentication, because an attacker with your administrator account will go looking for it. Retention must reach back further than an attacker could plausibly have been inside, which is months rather than days.

Set recovery time and recovery point objectives per system with the business owner, then write the measured time from your last test next to each one and treat the gap as the plan. Test by restoring from the awkward copy, not the convenient one. And try to delete a protected backup with your most privileged account, because that single test tells you whether any of the rest is real.

Frequently asked questions

What is the 3-2-1 backup rule?

Keep three copies of your data, on two different types of media, with one copy off site. It is a simple rule for surviving single points of failure, and it predates ransomware, which is why most organizations now extend it.

What is an immutable backup?

A backup that cannot be altered or deleted for a defined retention period, even by an administrator with valid credentials. It is enforced by the storage platform rather than by policy, which is what makes it useful against an attacker who has your admin account.

What are RTO and RPO?

Recovery time objective is how long you can tolerate a system being unavailable. Recovery point objective is how much recent data you can afford to lose. They are business decisions that determine backup frequency and recovery architecture, and they should be set per system.

Does cloud storage count as a backup?

Sync is not backup. A synced folder propagates deletion and encryption to every copy. Cloud platforms often include versioning and recycle bins that help, but retention is usually limited and an administrator can purge them, so treat them as one layer rather than the whole answer.

How often should we test restores?

Test individual file restores monthly, a full system restore at least quarterly for critical systems, and a full recovery scenario annually. Record how long each takes, because that measured time is your real recovery time, not the one in the plan.

Why do attackers target backups?

Because an organization that can restore does not pay. Deleting or encrypting backups is a standard step before deploying ransomware, which is why the backup system needs separate credentials, separate authentication and immutability.

Do we still need tape?

Not necessarily, but you need what tape provided: a copy that is physically or logically disconnected and cannot be reached from the production network. Immutable cloud storage with separate credentials achieves the same aim for most organizations.

What should we back up besides files?

Configuration and identity. Virtual machine images, database backups, directory and identity configuration, firewall and network device configurations, certificates and keys, email and collaboration data, and the documentation needed to rebuild. Data without configuration leaves you rebuilding from memory.

Is a backup enough to recover from ransomware?

It is necessary but not sufficient. You also need to remove the attacker's access first, rebuild to clean systems rather than restoring into a compromised environment, and be prepared for the fact that data stolen before encryption is not recovered by any restore.

Sources

  1. Small organisations guide to cyber security | Backing up your data, UK National Cyber Security Centre
  2. Data Backup Options, CISA
  3. SP 800-34 Rev. 1, Contingency Planning Guide for Federal Information Systems, NIST
  4. IR 8374, Ransomware Risk Management: A Cybersecurity Framework Profile, NIST
  5. SP 1800-11, Data Integrity: Recovering from Ransomware and Other Destructive Events, NIST
  6. #StopRansomware Guide, CISA
  7. Mitigating malware and ransomware attacks, UK National Cyber Security Centre
  8. Inhibit System Recovery, Technique T1490 - Enterprise, MITRE ATT&CK
  9. Data Encrypted for Impact, Technique T1486 - Enterprise, MITRE ATT&CK
  10. SP 800-209, Security Guidelines for Storage Infrastructure, NIST
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.