# VMware vCenter, security 0-day exploited since September

> CISA gives three days to patch vCenter after a zero-day exploited since September, with theft of the key that signs SAML tokens.
> By José Maldonado · Cybersecurity · 2026-11-21

Anyone administering VMware vCenter has three days to update, and the clock started running on Friday, November 20. And all because that day, the United States cybersecurity agency added CVE-2026-123890 to its catalog of exploited vulnerabilities, an authentication bypass rated 9.8 out of 10 in the VMware Directory Service, the component that underpins single sign-on across all of vSphere. The [CISA alert on vCenter dated November 20](https://www.cisa.gov/news-events/alerts/2026/11/20/cisa-adds-four-known-exploited-vulnerabilities-catalog) sets November 23 as the deadline for federal agencies and adds a requirement that does not appear in other advisories: checking whether the system was already compromised before applying the patch.

What makes this case special is not the score, it is the timeline. Mandiant attributes the exploitation to a group Google calls ARMORY CASTLE, previously tracked as UNC5221 and described as suspected of working for China, and places the first attacks in early September, against a vCenter that had already applied the fix for an earlier authentication bypass, the one from July. In other words, the attacker did not improvise: they kept a tool on hand waiting for the next flaw to open the same door. Broadcom published advisory VMSA-2026-0009 on November 19 and does not mention any exploitation; the public record arrives a day later, when CISA already has evidence of real-world use.

For an organization running vSphere in production, patching is not the only task. vCenter hands out the administrator passes for virtual machines, storage and the network, so whoever gets in there starts with the highest role and without needing to escalate privileges.

## A flaw that MFA does not stop

The flaw lives in \`vmdird\`, the directory service that stores users, groups and credentials for vSphere Single Sign-On. Broadcom describes it with the usual minimal formula: a person with network access to vCenter can bypass authentication. No account is needed, no password, and no user has to do anything; the [reconstruction by Google Threat Intelligence and Mandiant](https://techpastweek.com/es/blog/armory-castle-vcenter-cve-2026-123890-zero-day-patch) goes into the mechanism and explains why the 9.8 label is not just decoration.

The attack uses DCE/RPC, an old protocol for invoking functions on a remote machine, exposed on TCP port 2012, the one vCenter uses to govern Single Sign-On. There, the SRP-6a authentication mechanism relies on UUIDv1 identifiers generated from the system clock and therefore predictable. With that predictability, an attacker derives the shared session key without knowing any password and opens a session as \`administrator@vsphere.local\`, the most privileged account in the environment.

From that session, the path is short and highly profitable. The attacker can execute commands, download \`data.mdb\`, the directory service database, and extract the RSA private key with which the Security Token Service signs SAML tokens. Those tokens are the passes vCenter hands to anyone already authenticated, so forging a valid one for any user turns multi-factor authentication into a formality that is bypassed from above. The tool Mandiant associates with the campaign is called BONESAW and its documented functions include creating an SSO administrator account, copying the directory and reaching the ESXi hosts and the virtual machine disks.

The detail that turns a technical curiosity into a governance problem is the privilege asymmetry: an authentication flaw in an application server usually gives you an application, and this one hands over the hypervisor's master key.

## Three days, a weekend and two branches without a patch

CISA's deadline falls on a Monday, with a weekend in between, and that choice says a lot about the state of the flaw: the agency only adds to its catalog what is already being used in real attacks, and when it does so with three days' notice it is because it considers waiting worse than breaking a maintenance window. The list of fixed versions is the part of the advisory worth copying and taking to Monday's meeting.

Branch

Version with the fix

Note

vCenter 9.1.x

9.1.1.0200

Advisory VMSA-2026-0009 dated November 19

vCenter 9.0.x

9.0.2.0200

Updated on November 20

vCenter 8.0

8.0 U3m

Includes vSphere Foundation deployments

VMware Cloud Foundation 5.x

5.2.5

With vCenter 8.0 U3m, per KB 88287

Telco Cloud Platform and Infrastructure

Instructions in KB463116

Branches 3.0 to 5.2

vCenter 7.0, vSphere 6.5 and 6.7

No public patch

Extended support contract only

The advisory offers no temporary mitigation. In other cases Broadcom publishes a workaround while preparing the update; not here, and that leaves two paths: update or accept that the environment remains exposed. For anyone running vCenter 7.0 in production, the advisory offers no technical way out, only the commercial route of extended support, and that is the point where many small and mid-sized installations will discover that their version no longer receives fixes.

## What the vendor's advisory does not say

There is an asymmetry worth keeping in mind before citing sources. Broadcom's advisory does not mention exploitation at any point, and the NVD record did not include it on November 19 either. The evidence comes from those who investigated the incident and reported it: Google and Mandiant. A vendor not confirming an attack does not mean it does not exist, it means their advisory covers what they can demonstrate with their own records.

The second warning is about the figures. Google speaks of several incidents without saying how many, in which sectors or in which countries, so any estimate of victims is an invention. The third concerns the start date: the claim that the flaw was known before July rests partly on a timestamp that the report itself considers modifiable, and that is the kind of detail a headline usually rounds up.

None of the three warnings changes the urgency of the patch. They change what can be written rigorously about the scope of the attack, which is precisely the part that tends to get lost in the first wave of coverage.

The case does not arrive alone either. The same catalog update on November 20 included SAP NetWeaver CVE-2025-42884, Apache Roller CVE-2026-129448 and Cisco Expressway CVE-2026-76466, the latter already covered in the report on [the scripts the attacker rewrote with the help of AI](https://observatorioblockchain.com/en/cybersecurity/cisco-expressway-ataque-scripts-ia/). Four flaws, four three-day deadlines and the same message: the inventory of what runs in the data center is once again the most fragile part of security.

## How do you know if the attack has already been here?

CISA does not only ask for patching: it asks for forensic triage. In a vCenter, that means looking for what BONESAW leaves behind and that the update does not erase. The reasonable list includes SSO administrator accounts created outside the known change window, sessions started as \`\[email protected\]\` without a coherent authentication record, unusual downloads of the directory database and odd queries to port 2012.

It is also worth looking sideways. If the attacker reached the SAML signing key, they could forge valid tokens for any user and access other systems federated with that trust. A patch closes the door, it does not invalidate the passes already issued, so rotating that key and the associated certificates is part of the response, not an extra for the careful administrator.

One review worth adding that almost never appears on checklists: the appliance's own logs, because vCenter stores there the authentications and role changes that later allow the sequence to be reconstructed.

For organizations that prioritize work by CISA's catalog, there is a practical detail: the November 20 entry was not showing in the KEV when the first reports of the day were written. Anyone with an automated process that only watches that list can be a day behind an attack that has been underway for two and a half months.

## What will decide the outcome in January

The November 23 deadline does not close the case, only the emergency phase. After that, two open questions remain. The first is how many vCenter 7.0 installations are still alive without an extended support contract, because those have no fix available and their only way out is a migration that cannot be done in a weekend. The second is what Broadcom will do with the advisory: whether it updates it to mention the exploitation, as it has done in other campaigns, or whether it keeps silent and lets the evidence live only in a third party's report.

What can already be stated is that the combination of a predictable identifier in a legacy protocol and a token signing key in the same service is a pattern that repeats across the virtualization infrastructure of half the world. The [MPTCP flaw Linux left unpatched in RHEL](https://observatorioblockchain.com/en/cybersecurity/cve-2026-98432-mptcp-linux-sin-parche/) showed the same problem from another angle: the software nobody watches because it has been running for a decade is the one that decides whether the rest holds up.

The hypothesis left on the table is uncomfortable and verifiable in the next quarter: if the attacks began in September and nobody detected them until November, the number of compromised environments will be known from forensic investigations, not from press releases. Today's patch prevents the next one, it does not repair the previous one.
