Insights

AI Security

Storm-0940 used ARTEX to try more than 2,250 Microsoft 365 tenants after one vendor app fell

Microsoft says a China-state aligned actor used the agentic framework ARTEX to turn one compromised multi-tenant app into sign-in attempts against more than 2,250 tenants, in some cases reaching privileged cloud access in under three hours. Detection alone does not stop that tempo.

November 18, 2026 · 6 min read

Storm-0940 used ARTEX to try more than 2,250 Microsoft 365 tenants after one vendor app fell

After a vendor's multi-tenant application was compromised, Microsoft reports that Storm-0940, which it describes as a China-state aligned actor, used the open-source agentic pentesting framework ARTEX to attempt sign-in to more than 2,250 tenants. Microsoft assesses that Storm-0940 was responsible for the ConnectWise supply chain compromise, and says that in multiple instances the time from initial targeting to privileged cloud access was less than three hours.

That last number is the one I want defenders to sit with. The techniques were not new. Microsoft notes that OAuth application abuse is a typical part of Storm-0940's toolset, and that this time the actor used ARTEX to accelerate it to machine speed.

This is the first post in a short series on how threat actors are using agentic offensive tools right now. Each post stays at the level public reporting describes, and each one lands in the same three places: test the incident response plan, keep behavioral detections current, and be able to respond without waiting on a person.

What the reporting actually shows

Microsoft reports that as early as September 2026, Storm-0940 used ARTEX to automate attacks against hundreds of organizations in sectors including aerospace and defense, energy, financial services, government, semiconductors, software, and think tanks. Much of the access did not need a user to sign in. Microsoft describes the actor authenticating through applications that had already been granted access, then running reconnaissance, application credential changes, and privilege modification attempts through Microsoft Graph.

The password spraying around it was loud. Between October 10 and 14, Microsoft counted more than 151,000 failed logons against 327 accounts across eight organizations, peaking at 2,900 failures per minute on October 11, and Microsoft says that spray produced no successful authentication. The trusted application path is where it saw follow-on activity.

Wiz Research tracked an ARTEX-using actor that overlaps with Storm-0940 as it targeted IT and software vendors, focusing on identity providers (Okta, Entra, Cognito) and Microsoft 365 integrations. Wiz saw new IP addresses appear within minutes of a block, and the agent switch endpoints when a login page started rate limiting. GTIG tracks that campaign as UNC7228 and says it overlaps with Storm-0940.

Storm-0940 is not the only ARTEX user, and the attributions should stay separate. GTIG and Mandiant report that financially motivated extortion groups and additional suspected PRC-nexus clusters have also used ARTEX in incident response engagements. Those are other actors, not more Storm-0940.

Why the tempo matters more than the tool

GTIG found that individual ARTEX execution sequences ran in compressed bursts of roughly 2 to 25 seconds, and in one operation recorded 27 events in 23 seconds of active burst time. Microsoft spells out the consequence: this kind of activity generates alerts, but its speed leaves very little time for defenders to respond and remove the actor's access in a coordinated way.

GTIG is blunter. These frameworks currently trade stealth for velocity, and detecting an attack in time does little to stop it if nothing happens next. A SOC that sees the alert at minute two and pages someone at minute fifteen may be looking at an operation that is already done.

The tool is also a moving target. Elastic notes that the ARTEX author took the original repository down in October 2026 and public forks carried it on. Detections pinned to a project name, a default string, or a single IP address will age fast.

Test the plan against a trusted app gone bad

Most incident response plans I read assume the attacker comes in through a user. This one came in through an application the tenant already trusted. So the first tabletop question is simple: if a vendor tells you its multi-tenant app was compromised, who in your organization can disable that app's access, and how long does it take?

Microsoft's guidance for this incident was to review service principal sign-ins, investigate unexpected application credential or privilege changes, and keep affected application connections disabled until the vendor provides restoration instructions. Put that sequence in a tabletop inject and time it. Then add the detail that matters: the vendor notice arrives after the actor has already added an application credential.

A good exercise also asks what breaks when you disable the app. Here it was a cloud backup integration. If turning it off interrupts a recovery workflow, make that trade now, not during the incident.

Keep detections on behavior, not on ARTEX

Elastic's analysis of five agentic frameworks makes the point I would build on: every action these tools take still lands as a process, a web request, an API call, or a sign-in, so detections anchored on behavior outlast detections anchored on a source address, a User-Agent, or a framework fingerprint.

For an intrusion shaped like this one, that means watching identities, not tools:

  • Service principal sign-ins from new infrastructure, especially for third-party apps.
  • A new credential added to an application and a new privileged role assignment in the same window.
  • Bursts of Graph reconnaissance from one app identity.
  • Spray patterns across many accounts, even when every attempt fails.

GTIG's mitigation guidance adds one I would not skip: keep enough identity, service principal, and control-plane telemetry history to see a sequence, because a burst only means something next to what came before it.

Respond at the speed of the attack

The third piece is the one most teams have not built. Elastic recommends deciding in advance which high-confidence detections start containment automatically, and pairing any block on a source address with revoking the identity's sessions and tokens, because these frameworks come back from new addresses.

Elastic also warns that revocation is not instant. In Entra ID, Elastic explains, an access token issued before a service principal is disabled generally stays valid until it expires, typically 60 to 90 minutes, unless continuous access evaluation applies. An agent that works in seconds can do a lot inside that gap.

So the autonomous response step is not "page the on-call." It is a pre-approved action, such as disabling a third-party service principal when it adds a credential and grants itself a privileged role, followed by a check that the next request actually fails. A person reviews the decision afterward, and that review is where you learn whether the trigger was right.

If your playbook for a compromised vendor app still starts with a ticket, run the scenario once with a timer and see where the first three hours go.