Cloud backups fail during attacks when the same credentials can delete them
An InfoWorld contributor argues that cloud versioning, replication and snapshots keep services running but do not guarantee recovery after an attacker with stolen credentials. Microsoft's September 2026 Azure incident is cited as an example.
TL;DR
- Versioning, replication and snapshots are built for availability, so they can fail when an attacker, not an outage, is the cause.
- Recovery copies must be out of reach of production credentials, and locked where the platform allows it.
- Recovery tests should restore the data and confirm the application works, not only redeploy infrastructure.
What happened
An InfoWorld opinion piece by Alka Tandon, GM of Cloud at Rubrik, starts from an incident Microsoft described in September 2026. Attackers used compromised Azure identities to delete cloud resources and go after recovery protections. In a destructive sequence lasting about seven minutes, they attempted more than 100 deletions of storage accounts and deleted most of them. The author uses this case to show how much damage one identity with enough privileges can do.
The article notes that teams usually fall into one of two camps. Some spend heavily on disaster recovery because they fear a hyperscaler outage. Others assume data in AWS or Azure is already safe. The author says both groups misread what native resilience tools are for, since those tools behave differently when the cause is an attacker.
Where the safety net breaks
The piece walks through several common assumptions. The first is that S3 versioning equals a backup. Versioning helps with accidental overwrites, but it does little when an attacker holds valid credentials. The article cites a January 2025 Halcyon report on the Codefinger ransomware campaign. In that case, attackers used compromised AWS credentials and SSE-C server-side encryption to make data unreadable, then threatened deletion through S3 life-cycle policies. The advice is to limit who can change those policies and to use Object Lock or MFA Delete.
The second assumption is that replication protects data. The author says replication is designed mainly for availability. Corrupted or encrypted data can be copied to other regions along with good changes, and deletion behavior differs between services. Critical recovery copies should sit in a separate account that production credentials cannot reach.
The third is that snapshots are enough. The article points to Code Spaces, which in 2014 lost data, backups, machine configurations and snapshots after an attacker reached its AWS control panel. The company shut down within days. The recommended fix is to lock snapshots so nobody can delete them before retention ends, including admins.
Infrastructure as code and stateless apps
Infrastructure as code makes environments quick to recreate, but it does not restore data. Terraform can rebuild a database server, while customer records and transaction history still need a separate recovery path. The author suggests recovery tests that restore the data an app depends on and then verify the restored system works.
The article also challenges the idea that stateless applications hold no data worth protecting. Such apps often write to Postgres, Amazon S3, DynamoDB, Redis, Snowflake or Apache Kafka. Teams should map every service an application writes to and assign someone to protect each one. Audit logs get special attention: security teams rely on them to find out what an attacker did and whether restoring is safe, so they belong in retention and recovery planning.
Why it matters
For developers and platform teams, the core message is that backup design has to assume the attacker has production credentials. Two questions apply to any native tool: can someone with stolen credentials delete the recovery copies, and has the team ever restored from them? The author also notes that executives often ask when the business will be back online, while the engineers doing the work may not yet know how much is damaged. Knowing what your cloud tools can and cannot recover before an incident is the practical takeaway.
Sources
This article was written by Blozhik from the reporting below. Read the originals for full detail.
InSix data protection myths cloud-native teams need to stop believingInfoWorld