When the Blast Door Was a GitHub Comment

September 29, 2026 · SPR{K3 Research

Storm-3168 walked into an Azure tenant with credentials someone published in a public GitHub issue, deleted 100 storage accounts in seven minutes, and Microsoft's own writeup is that only the resource locks set beforehand survived

On September 25, Microsoft Threat Intelligence published a writeup of an intrusion it is now calling "the first documented agentic ransomware operation." The actor is tracked as Storm-3168 and, per Microsoft, is linked to JADEPUFFER — the same operator the digest has already logged twice this year: once in July 2026 for the Langflow CVE-2025-3248 initial-access campaign, and again on July 21 for the ENCFORGE ransomware purpose-built to encrypt AI model weights, vector databases, and training corpora. Same actor, third substrate in three months.

The tier-1 wrap-up landed on September 28 at The Hacker News, The Register, Security Affairs, and Techtimes, whose headline reads: "AI Ransomware Wiped 100 Azure Accounts in 7 Minutes: Only Pre-Configured Locks Survived."

How the door opened

The initial access was not a zero-day. It was two Azure service-principal secrets — client ID, secret, tenant ID — that an employee at the victim organization had pasted, in plaintext, into a public GitHub issue. Per Microsoft's own writeup: the visible content of the issue was later removed. The secret, however, remained retrievable through GitHub's per-issue edit history — a channel most secret-scanning pipelines don't consult and most repo owners can't audit.

The attacker's user agent, per Microsoft, was Python-requests/2.34.2.

The tempo, minute by minute

Microsoft reconstructs the campaign in early June 2026 (dates in Skycloak's analysis and Cyberpress's coverage):

On the Storage-Account side, the deletions that hit accounts protected by pre-configured Azure resource locks or storage-account-level deletion protection were blocked. On the Azure SQL side, the actor tried to delete databases via an API version Azure no longer supported and got API errors instead of deletions. That is the entire list of things that stood between the actor and total tenant erasure inside the 35-minute window.

What actually held the line

Read Microsoft's own recommendations in order. First it names the surviving control: pre-configured resource locks and deletion protection. Everything else — Defender for Cloud plans across Resource Manager, Storage, Key Vault, App Service, and Databases; least privilege on service principals; treating publicly-exposed credentials as compromised regardless of removal; scaling investigation and response with Project Perception and MDASH — is Microsoft's post-incident answer for the next attack. Nothing in that list runs at the seven-minute-per-100-accounts tempo the actor was operating at, unless it was already switched on.

Techtimes' headline is the accurate one. The classical detect-triage-contain loop does not fit inside a seven-minute burst on a hyperscaler's control plane. What survives is what was pre-authorized to survive.

The pattern this campaign closes

Two weeks ago Cisco Talos disclosed CLOSEDQUORUM, a Windows malware family that polls a panel of four commercial LLM endpoints (DeepSeek, Qwen, Mistral, Gemini), takes a vote, and executes the winning action. Two weeks before that, GreyNoise disclosed that a Russian-speaking operator running a Codex + DeepSeek fleet compromised 11 PaperCut organizations inside a 26-second window. Both are the endpoint-side shape of "attacker-agent-fleet at machine tempo."

Storm-3168 is the same architecture, moved one tier up. The fleet is the vendor's own service-principal abstraction. The C2 channel is the hyperscaler's own resource-manipulation API. The Discord webhook is replaced by ListKeys. The infrastructure the attacker deletes belongs to a Fortune-scale enterprise that owns the tenant, not a print server on a university LAN.

Same actor lineage — JADEPUFFER encrypting model weights in July, Langflow initial-access before that, Storm-3168 destroying cloud control-plane state now. In three months, one actor has moved from a foothold in an AI orchestration platform, to encrypting the artifacts that platform trained, to deleting the cloud accounts those artifacts lived in.

Two things for defenders to notice

Service-principal secrets are agentic ammunition, and they live in more places than your code. GitHub issue edit history is one such place. Team wiki drafts, deleted-but-cached Slack messages, screenshots pasted into Jira, terraform state files pushed to the wrong branch — the Skycloak wrap-up makes the wider point that static, high-value, long-lived NHI (non-human identity) credentials are the primitive most likely to survive both the initial mistake and the subsequent cleanup. Microsoft's own recommendation — "treat credentials that have been publicly exposed as compromised" — is the rule that assumes cleanup does not undo exposure. The corollary is that any control that depends on "we deleted the leaked secret" is not the control that will stop the next Storm-3168.

Pre-authorization is a runtime control now. The classical framing is that resource locks and deletion protection are compliance items — you set them because governance asked. Storm-3168 recasts them: they are the only class of control that runs at agentic-attack tempo, because they answered the question of whether the delete was allowed before the delete arrived. On the platform side, this is why Microsoft's Sept 25 recommendation stack ends at agentic defenses like Project Perception — the classical human-in-the-loop response cadence has fallen behind the attack. On the customer side, the takeaway is smaller and more immediate: enumerate the resources whose destruction would be irreversible, and check today whether they carry a lock. What Techtimes' headline says is empirically what happened. On a seven-minute burst, the accounts with locks are the accounts you have left.

The next Storm-3168 will not need a GitHub-issue leak. It will find its own credential path — a fork of the same edit-history trick, a stolen terraform state file, an OAuth token in a browser cache. The attacker fleet is now the vendor's own control plane. The blast door is whatever you already set before the fleet arrived.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.