Cloud ransomware does not need to encrypt anything if the attacker can delete your backups and snapshots with the same administrator role that manages them. Microsoft's 2025 reporting on Storm-0501 described a group that took over cloud administrator identities, removed resource locks and immutability policies, deleted recovery points and then demanded payment. All three major providers now offer native controls that survive a compromised administrator: AWS Backup Vault Lock and logically air-gapped vaults, Azure Backup locked immutable vaults with multi-user authorization, and Google Cloud Backup and DR backup vaults with enforced, locked retention. This guide compares them control by control, explains the identity design that makes them effective, and covers enforcement at scale, monitoring and restore testing. It is for cloud platform, backup and security teams, particularly in multi-cloud estates. Rebuilding after an attack is covered in the Cloud Recovery and Clean Room Guide; the general principles are in the Backup Security and Immutable Backup Guide.
Key takeaways
- Only the irreversible setting counts. AWS compliance mode after its grace time, Azure "Enabled and locked" and a Google retention lock past its effective date are the states an attacker with full administrator rights cannot undo. Governance mode and unlocked settings are tests, not defences.
- Put the second key outside the blast radius. AWS multi-party approval uses a separate recovery organisation; Azure Resource Guard can sit in another subscription or tenant. A second approver in the same tenant as the attacker adds little.
- Backups are a non-human identity problem. Backup service roles, restore roles, vault managed identities and automation pipelines hold the permissions that matter; scope them and watch them.
- Locked retention commits cost. Every provider charges for locked recovery points until they expire, so set retention deliberately before locking.
- A backup you have not restored is a hypothesis. Schedule automated restore tests and record restore times against recovery time objectives.
The threat: administrator-level destruction of recovery data
The attack pattern is consistent across clouds. The actor obtains a privileged cloud identity (a stolen administrator session, a federated trust, synchronisation credentials or a long-lived access key), escalates to an owner-equivalent role, and then targets recovery before or instead of encryption. MITRE ATT&CK describes the impact steps as T1490 Inhibit System Recovery, T1485 Data Destruction and T1486 Data Encrypted for Impact, with T1078.004 Cloud Accounts and T1098.003 Additional Cloud Roles on the way in.
- Recovery point deletion: deleting backups, snapshots and restore point collections through normal APIs.
- Retention manipulation: shortening retention or swapping backup policies so recovery points age out quietly.
- Protection removal: deleting resource locks, disabling soft delete, removing immutability policies, or deleting the vault's encryption key so backups become unreadable.
- Account-level destruction: closing or detaching an account. AWS documents that if an account containing a vault is closed and not reopened within 90 days, the vault contents are deleted even if a vault lock was in place.
Native backup controls are designed for this attacker: someone who is, from the platform's point of view, a legitimate administrator. See the Cloud Ransomware Guide.
The three native services compared
| Control | AWS Backup | Azure Backup | Google Cloud Backup and DR |
|---|---|---|---|
| Vault type | Backup vault; logically air-gapped vault | Recovery Services vault; Backup vault | Backup vault |
| Reversible immutability (test phase) | Vault Lock in governance mode: removable by users with sufficient IAM permissions | Immutable vault "Enabled": can be disabled | Minimum enforced retention before the lock effective date: can be raised or lowered |
| Irreversible immutability | Vault Lock in compliance mode after a grace time of at least 72 hours (ChangeableForDays 3 to 36,500); deletes and lifecycle changes denied, including for the root user | Immutable vault "Enabled and locked", with WORM storage | Retention lock past its effective date: retention can only be increased, and Google states not even a Project Owner can decrease it |
| Retention bounds | MinRetentionDays (at least 1) and MaxRetentionDays (up to 36,500) | Immutability based on policy retention, or a fixed duration | Minimum enforced retention from 1 day to 99 years, or inherit from the backup rule |
| Deletion safety net | The vault lock itself; deletion is denied while it is active | Soft delete enforced by default, 14 days (extendable to 180) | Enforced retention prevents manual deletion by anyone, including Google |
| Second-person control | Multi-party approval through AWS Organizations, for logically air-gapped vaults | Multi-user authorization with Resource Guard | No equivalent documented on the vault page; rely on IAM separation and the retention lock |
| Isolation from the source account | Logically air-gapped vault stores data in a service-owned account and can be shared with individual account IDs through AWS RAM for restore | Vault in a separate subscription; Resource Guard in another subscription or tenant | Vault access restriction (current organisation, current project, organisation with backup appliances, or unrestricted), fixed at creation; centralised vault project |
| Encryption | AWS owned key (default for air-gapped vaults) or customer managed KMS key, chosen at creation | Platform-managed key, or customer-managed key in a same-region Key Vault with purge protection | Not covered on the vault page reviewed; check the Backup and DR encryption documentation |
| Compliance assessment | Vault Lock assessed by Cohasset Associates for SEC 17a-4, CFTC and FINRA environments | Not stated in the pages reviewed | Not stated in the pages reviewed |
Provider details for each service are in the AWS Backup Security Guide and the Azure Backup and Recovery Services Vault Security Guide.
Choosing a design
- Single cloud, single account or subscription: lock immutability in the main vault and keep the second-person control (MUA on Azure). This defeats deletion by a compromised administrator but not loss of the whole account, tenant or billing relationship.
- Separate backup account or project: copy backups into a vault in a dedicated account, subscription or project with different administrators. This is the minimum for regulated or critical workloads. On AWS the logically air-gapped vault provides this without managing a separate data account.
- Separate organisation or tenant for the recovery path: AWS multi-party approval expects a recovery organisation with its own IAM Identity Center; Azure Resource Guard can live in another tenant. Use this where a full identity provider compromise is in your threat model.
- Cross-cloud copies: some organisations copy critical data to a second provider or an off-cloud immutable store. Native services do not do this for you, and it adds a second identity estate to protect. Weigh it against concentration risk. See the Cloud Concentration Risk and Exit Strategy Guide.
Where identity fits: backup roles and service identities
Immutability blocks deletion, but most other backup risks are identity risks: who can change plans, who can restore data elsewhere, and which non-human identities can read everything. See the Backup Identity Isolation Guide.
| Identity | AWS | Azure | Google Cloud | Control |
|---|---|---|---|---|
| Human backup administrators | IAM roles via IAM Identity Center | Backup Contributor, Backup Operator | Backup and DR IAM roles in the vault project | Just-in-time access with approval; no standing owner-level roles on backup accounts |
| Second-person approvers | Multi-party approval team in a recovery organisation | Security admin owning Resource Guard (Backup MUA Admin) | n/a | Different people and, ideally, a different identity provider from the backup admins |
| Backup service identity | Backup service role and service-linked roles | Vault managed identity | Backup and DR service agents | Read access to sources only; never interactive |
| Restore identity | Restore role with AWSBackupServiceRolePolicyForRestores permissions | Backup Operator plus rights on the target resource group | Restore permissions on target project | Restoring to a new location is a data exfiltration path: alert on restores to unexpected accounts |
| Automation | Pipelines deploying backup plans and Organizations backup policies | Pipelines deploying vaults and Azure Policy | Pipelines deploying backup plans | Workload identity federation; scoped to backup resources |
- Keys in personal vaults: the LastPass breach shows the failure mode: AWS access keys and decryption keys stored in an engineer's password vault let attackers copy customer vault backups from cloud storage. Backup access should come from federated roles, not keys that a person holds.
- Encryption keys are part of the backup: if a customer-managed key protects backups, the key's administrators can make every backup unreadable. Keep key administration separate from backup administration, enable deletion protection, and back up key material according to the provider's model. See the Cryptographic Key Management Guide.
Enforcing coverage at scale
Protection that depends on each team remembering to configure it will have gaps. Enforce backup and immutability centrally, and report on exceptions.
- Tag-driven backup plans: define backup tiers (for example gold, silver, none) with retention and copy rules, and attach plans by tag. On AWS, AWS Organizations backup policies can push plans to accounts; on Azure, Azure Policy can configure backup on tagged resources; on Google Cloud, use backup plans in a centralised vault project.
- Preventive guardrails: deny deletion of vaults, vault locks, Resource Guards and backup-related key material except from a break-glass role, using AWS service control policies, Azure Policy and deny assignments, or Google Cloud organisation policy and IAM deny policies. See the AWS Organizations and SCPs Guide and the Azure Policy and Management Groups Guide.
- Pilot reversible settings: run AWS governance mode, Azure "Enabled" immutability or an unlocked Google minimum retention for a few weeks, checking cost, policy changes and job success.
- Lock with intent: move to compliance mode, "Enabled and locked" or a locked retention. Record the retention values and who approved them; there is no rollback.
- Isolate a copy: configure copies to a separate account, subscription or project (or an AWS logically air-gapped vault) for top-tier data.
- Report: coverage by tier, lock state per vault and exceptions with owners and expiry dates.
Retention, cost and organisation size
Locking turns retention into a commitment. Each provider charges for locked recovery points until they expire: AWS notes that locked backups cannot be deleted until their lifecycle completes, Azure charges soft delete retention beyond 14 days, and Google states that backups under enforced retention incur storage charges for the full duration. Decide retention from three inputs before locking anything.
| Input | Question | Effect on settings |
|---|---|---|
| Detection window | How long could an attacker be present before you notice? | Immutable retention for the isolated copy should exceed it, with margin |
| Regulatory and contractual retention | What must be kept, and what must be deleted on schedule? | Maximum retention bounds (AWS MaxRetentionDays), separate vaults for data with deletion obligations. See the Backup and Recovery Policy Template |
| Recovery objectives | Which point in time do you need, and how fast? | Backup frequency, copy schedule, cross-region copies and restore capacity |
- Do not lock "keep forever": AWS documents that recovery points with retention set to "Always" are kept permanently once a compliance-mode grace time expires. Use explicit durations.
- Use bounds to stop mistakes and manipulation: minimum retention on the vault means a changed backup plan cannot create short-lived recovery points that age out before anyone notices.
- Personal data: long immutable retention conflicts with erasure obligations for some data. Keep such data in vaults with shorter, documented retention and record the decision.
- Small organisation: one locked vault per cloud with a separate administrator, soft delete or minimum retention at a level the budget allows, and a quarterly restore test is a realistic baseline. If there is no second person for approval, a trusted external provider or senior manager can hold the second key.
- Large or regulated estate: tiered plans pushed by policy, isolated copies in dedicated accounts or projects, a recovery organisation or tenant for approvals, automated restore testing and central reporting of lock state and coverage.
Detection and measurement
| Signal | AWS | Azure | Google Cloud |
|---|---|---|---|
| Lock or immutability change | CloudTrail PutBackupVaultLockConfiguration, DeleteBackupVaultLockConfiguration | Activity log; built-in alerts for disabling soft delete | Cloud Audit Logs Admin Activity on the backup vault |
| Recovery point deletion | CloudTrail DeleteRecoveryPoint | Built-in Azure Monitor security alerts for deletion of backup data | Admin Activity and Data Access logs |
| Plan or retention changes | CloudTrail UpdateBackupPlan, DeleteBackupPlan | Activity log policy writes | Admin Activity on backup plans |
| Privilege escalation | IAM policy attachments and role creation in backup accounts | Role assignments, Global Administrator elevate access | IAM policy changes on vault projects |
| Restore to unexpected location | CloudTrail StartRestoreJob with unusual target account or Region | Restore jobs to unexpected subscriptions | Restore operations to unexpected projects |
| Job failures at scale | AWS Backup job events in EventBridge | Built-in alerts for backup failures | Backup and DR job monitoring |
Send these to a SIEM that lives outside the accounts being protected, so the attacker who reaches backup administration does not also control the alerting. See the Cloud Logging and Audit Trail Guide.
Restore testing
- AWS Backup restore testing: scheduled restore testing plans pick the latest or a random recovery point, restore it, keep the restored resource for 1 to 168 hours for optional validation, then delete it. Restored resources carry the tag
awsbackup-restore-test. The AWS Backup Audit Manager control "Restore time for resources meet target" checks restore duration against your target. - Azure: restore into an isolated subscription on a schedule, and use Cross Region Restore drills for critical virtual machines.
- Google Cloud: restore from the backup vault into a test project on a schedule and record time to usable service.
- Validate content, not just completion: check that the application starts and data is consistent, and scan restored data for malware before it rejoins production (NIST CSF 2.0 RC.RP-03).
Metrics
- Percentage of tier-one workloads with a copy in an irreversibly locked vault.
- Percentage of tier-one workloads with a copy isolated in a separate account, subscription, project or air-gapped vault.
- Restore test success rate and median restore time against the recovery time objective.
- Number of standing owner-level assignments on backup accounts and projects (target: zero).
- Time from a backup security alert to triage.
Framework mapping
| Area | NIST CSF 2.0 | NIST SP 800-53 Rev. 5 | Microsoft cloud security benchmark v1 | MITRE ATT&CK addressed |
|---|---|---|---|---|
| Automated backups by policy | PR.DS-11 | CP-9 | BR-1 | T1485 |
| Immutability and isolation | PR.DS-11 | CP-9 | BR-2 | T1490 |
| Separation of backup and security duties | PR.AA-05 | AC-5 | BR-2 | T1078.004, T1098.003 |
| Monitoring backups | DE.CM-09 | CP-9 | BR-3 | T1490 |
| Restore testing and verification | RC.RP-03 | CP-4, CP-10 | BR-4 | T1486 |
Common failure patterns
- Governance mode or unlocked immutability in production: reported as "immutable" but removable by the administrator the attacker has become.
- Backups in the same account as production with the same administrators: one identity compromise covers both.
- Second-person control inside the compromised boundary: an Azure Resource Guard owned by the backup team, or approvers who all sign in through the same compromised identity provider.
- Snapshots treated as backups: instant restore snapshots and disk snapshots live with the resource and its permissions.
- Locking without a retention decision: recovery points kept for longer than needed at locked cost, or too short to outlast attacker dwell time.
- Restore never exercised: missing permissions, network dependencies or keys found during the incident instead of in a test.
Lessons from incidents
- Storm-0501 (Microsoft, August 2025): after compromising a Global Administrator, the group elevated to User Access Administrator and Owner across Azure subscriptions, deleted resource locks and blob immutability policies where they blocked deletion, and targeted snapshots, restore points and Recovery Services vault data. Where data stayed protected it tried to encrypt with a new Key Vault key and delete the key; Key Vault soft delete allowed recovery. Lessons: locks and reversible policies do not stop an Owner, irreversible vault immutability and purge protection do. See the Storm-0501 Threat Actor Profile Guide.
- LastPass (2022): the attacker used access keys and decryption keys from a DevOps engineer's vault to copy backups from cloud storage. Lesson: protect read access and key separation, not just deletion; immutability does nothing against copying.
Practitioner checklist
- Classify workloads into backup tiers and attach plans by tag through central policy.
- Move tier-one vaults to the irreversible state: AWS compliance mode, Azure "Enabled and locked", Google locked retention.
- Keep an isolated copy in a separate account, subscription, project or AWS logically air-gapped vault.
- Place the second-person control (AWS multi-party approval, Azure Resource Guard) outside the production tenant or organisation.
- Remove standing owner-level roles from backup accounts and require approval for backup administration.
- Separate encryption key administration from backup administration and enable key deletion protection.
- Alert on lock changes, recovery point deletion, retention changes and restores to unexpected locations, in a SIEM outside the protected accounts.
- Run scheduled restore tests and track restore time against recovery time objectives.
- Review locked retention values and costs before each lock, and annually after.
Standards and references
- AWS: AWS Backup Vault Lock
- AWS: Logically air-gapped vault
- AWS: Multi-party approval for logically air-gapped vaults
- AWS: Restore testing
- Microsoft Learn: Immutable vault for Azure Backup
- Microsoft Learn: Multiuser authorization using Resource Guard
- Google Cloud: Backup vaults in Backup and DR
- Microsoft Security: Storm-0501's evolving techniques lead to cloud-based ransomware
Related NHI Mgmt Group resources: Cloud Recovery and Clean Room Guide · Backup Security and Immutable Backup Guide · Backup and Cyber Recovery Buyer's Guide · Ransomware Response and Recovery Guide