- Microsoft reports that in one Azure tenant, attackers using two hijacked application logins tried over 100 storage account deletions in about 7 minutes. Most succeeded, but resource locks and deletion protection blocked a few.
- This matters because safeguards that sit apart from a login keep working even when that login has broad rights. Microsoft's evidence covers a few accounts in one tenant, not locks in general.
- Check which application logins hold broad roles like Contributor, turn on locks and deletion protection for storage and backups, and revoke or rotate any secret posted publicly.
Automated attacks are fast. That is the part people remember. The more useful detail is what slowed this one down. In one Azure environment, a few storage accounts survived. The safeguards on them did not depend on which login was used.
What Microsoft found
Microsoft Security Research tracks the actor as Storm-3168. Others call it JadePuffer. Sysdig discovered the group in July 2026. Reports describe it as the first documented ransomware operation run by AI agents. An AI agent is software that carries out tasks on its own.
Microsoft's report covers one tenant, meaning one organization's Azure account. It calls this the first detailed look at the group's Azure activity. It does not say how many other organizations were targeted.
Two service principals were compromised in that tenant. A service principal is an identity that lets an application sign in to Azure. One did the scouting. The other did discovery, deletion and credential collection.
The scout read Azure resources for about 15.5 hours. It made more than 300 successful reads. About 90 minutes after it started, the second identity listed virtual machines and resource groups across two subscriptions. That took five seconds. Microsoft says the timing and split of work "strongly indicates automated or scripted execution."
The attacker used rights it already had
No clever exploit was needed at this stage. Microsoft says each action matched a role the identity already held. A role granted through a group allowed the storage deletions. Direct Contributor access allowed the deletion of a Key Vault, a Function App and an App Service plan.
Here a technical issue becomes a people issue. The report does not say how the roles were assigned. It does show what a broad role is worth to whoever takes over the identity. The attacker gets every right the application had.
The attacker also targeted recovery. It tried to delete Azure Site Recovery locks and an Azure Backup protection lock. Those attempts failed. About 30 minutes after the last deletion, it sent more than 30 successful requests for storage account access keys.
What held, and what only looked like a defence
Most of the targeted storage accounts were deleted. For a few, the requests were blocked. Those accounts had resource locks or account-level deletion protection.
Microsoft draws a clear point from this. A safeguard that sits apart from the identity keeps working, even when that identity has wide rights. The evidence covers a few accounts in one tenant. It does not show that locks stop such attacks in general.
Not every failure was a defence. Every attempt to delete Azure SQL databases failed. The attacker used an API version that the database resource type does not support. That was the attacker's mistake, not a control. It is not something to rely on.
The lesson is simple. Suppose one login can create a resource and also remove its protection. Then the protection adds little. Safeguards work best when a different set of permissions guards them.
A secret that outlived its deletion
Microsoft could not tell how the attacker first got in. It did find something else. An employee of the affected organization had posted a service principal's client ID, client secret and tenant ID in a public GitHub issue. The issue was later edited. The old text stayed visible in the issue's public edit history.
Microsoft could not confirm that this secret was used in the activity it describes. The finding still carries a rule. Editing or deleting a post does not cancel the credential. Treat anything exposed in public as compromised. Then revoke or rotate it.
What the report does not show
Microsoft saw no ransom note. It did not confirm that any data was stolen. It says the pattern fits tactics that can support ransomware and extortion. That is a careful reading, and it should stay one.
Questions to put to your team
Ask which service principals in your cloud accounts hold Contributor or similar broad roles. Ask whether each one needs those rights for its job.
Ask whether resource locks and deletion protection are on for storage, backups and recovery vaults. Ask who can remove them. Ask whether removal raises an alert.
Ask what happens when someone pastes a secret into a public issue or repository. The answer should include revoking or rotating it, not only editing the post. Microsoft also suggests favouring credentials that do not last long.
The attacker moved at machine speed. The safeguards that worked were set in advance, by people, before any of it began.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Microsoft Security.





