Skip to main content
Back to Blog
Enterprise Security

Privileged Access in Under Three Hours, and Nobody Logged In

On October 15, 2026, Storm-0940's agents took over ConnectWise Cloud Backup application identities and walked into customer tenants without a single interactive sign-in. In 2027 that is the default adversary, and the only control that survives it re-checks authority after the first ALLOW.

Mark Rogge6 min read
Under Three Hours. No Login. — EnforceAuth 2027 series, no. 02 of 18

ENFORCEAUTH 2027 — NO. 02 OF 18 — ATTACKERS

By Mark Rogge, CEO of EnforceAuth | 6 min read | for identity and detection engineers

Between October 10 and October 14, 2026, a China state-aligned group Microsoft tracks as Storm-0940 ran a password spray against 327 accounts at eight organizations. 151,000 failed logons. A peak of 2,900 attempts per minute. Zero successful authentications. By most dashboards, the defenders won.

On October 15, the same operators took control of ConnectWise Cloud Backup application identities and authenticated into customer environments. No interactive user sign-in was required. Sign-in attempts followed against more than 2,250 tenants. In some of them, time to privileged access was under three hours. That is from Microsoft's October 21 report, with additional detail from CSO Online.

The spray was the decoy. The application identity was the door.

What ARTEX actually is

ARTEX is a Go-based agentic penetration-testing framework. Configurations recovered in the campaign pointed at DeepSeek, GLM and Grok models. It handles reconnaissance, password spray and OAuth application abuse as tasks, not as a human operator's keystrokes. CrowdStrike has separately reported ARTEX in use against Korean financial firms, and Trend Micro reported on October 27 that a single attack platform ran 44 AI agents against 19 or more organizations, 16 of them in Japan.

Three operational properties matter for 2027:

  • Sessions ran 24 to 30 hours. A human red teamer sleeps. The agent did not.
  • New source IPs appeared within minutes of blocks. IP-based response was outrun on every cycle.
  • Persistence was authorization, not malware. The agents added credentials to existing OAuth applications and changed privileged role assignments. Nothing dropped to disk. Everything was a permitted API call.
The attacker never needed your password. It needed one application identity that your tenant already trusted, and a policy that never asked a second question.

Why the first ALLOW is the whole game

Most authorization in production is evaluated once, at token issuance, and then trusted for the life of the session. An application identity with a valid client secret or certificate gets a token; the token carries scopes; the scopes are honored until expiry. That design assumed the entity holding the token was the entity the grant was issued to, acting for the purpose the grant was issued for.

Storm-0940 broke both assumptions without breaking any cryptography. The credential was real. The grant was real. The purpose had changed, and nothing in the path was built to notice.

The counter-argument is that conditional access and identity threat detection already cover this. They cover part of it: anomalous sign-in location, impossible travel, token replay. But in the ConnectWise case there was no sign-in to evaluate. The action that mattered was "add a credential to this OAuth app" and "assign this role," and those are resource-level operations, not identity-level events. A control that only looks at the login never sees them.

The 2027 forecast

By the end of 2027 I expect agentic frameworks like ARTEX to be the default tooling for both state and criminal operators against enterprise tenants, with application and workload identities as the primary initial-access target rather than human accounts. Gartner forecasts that AI agents will halve the time to exploit account exposures by 2027. Okta's September 2026 outlook puts identity-chained pivots from human to non-human identity as the primary vector. The Storm-0940 timeline, from spray to privileged access in days and from foothold to privilege in under three hours, is the floor, not the ceiling.

The implication for defenders: your mean time to detect is being measured against an adversary that completes its objective inside a single 24-hour session. If the check happens once at the front door, the door is the only thing you will ever log.

What continuous re-evaluation looks like

Here is how we build for it at EnforceAuth, stated as what ships and what does not.

Policy on the action, not the token. Our policy decision point (PDP), built on OPA and Rego, evaluates each request against the resource: "may this identity add a credential to this OAuth application right now, under this grant?" It is a separate question from "is this token valid," and it is answered at the resource, in path.

Session-scoped authority with expiry. A grant for a backup workload carries the actions a backup workload performs. Role assignment and credential management are not among them. Policy denies by default outside the scope, and the scope expires on a schedule shorter than a 24-hour attacker session.

OAuth app and role changes as policy inputs. EvidenceIQ, part of the EnforceAuth platform, maintains a relationships graph across identities, grants, applications and resources. When a credential is added to an application or a privileged role changes, that is a graph event, and policies that depend on the prior state are re-evaluated. An ALLOW issued at 09:00 does not survive a privilege change at 09:40 unexamined.

Evidence of every decision. Each decision is logged, tied to the exact signed policy version, and replayable. When the question after an incident is "which of the 2,250 tenants did the identity actually touch, and what did it do there," the answer is a query against a tamper-evident record, not a reconstruction from six log sources.

Our latency goals for the PDP (p50 under 2ms, p99 under 5ms) are design targets, and I will keep saying so until we publish measurements. Verdict, our agentic firewall, is coming and is not in this story. Everything above it is shipping today across multi-tenant SaaS, single-tenant SaaS, BYOC and air-gapped deployments.

The Storm-0940 operators did not log in. In 2027, neither will anyone else worth worrying about. Build the check where the action lands.

EnforceAuth is the runtime authorization layer for AI agents and non-human identities, built on OPA and Rego. See enforceauth.com and read the EvidenceIQ announcement.

About EnforceAuth

EnforceAuth is the OPA/Rego authorization control plane. The customer's OPA decides what people, services, and AI agents may do at action time. Verdict covers high-impact agent actions. EvidenceIQ keeps the record behind every decision.

Follow us on LinkedIn