A cut-paper illustration of a tall teal server rack with its side door standing open, a paper key hanging beside it, and a dotted line reaching in from outside

On November 19, Broadcom published an emergency patch for VMware vCenter with a CVSS score of 9.8. The flaw, CVE-2026-123890, lets anyone with network access to a vCenter server walk away with an authenticated session as administrator@vsphere.local. No password. No phishing email. No stolen credential. The attacker just talks to the directory service on the right port, and the directory service hands over the keys.

Google's threat intelligence team says a suspected China-linked espionage group had already been using it since early September. The patch landed ten weeks after the exploitation started. That gap is the story, and it is the part that should change how you think about the next one.

What the flaw actually is

vCenter is the control plane for a VMware estate. It configures, controls, and monitors the ESXi hosts and every virtual machine on them. Inside vCenter sits the VMware Directory Service, vmdird, which handles identity for Single Sign-On. One of its interfaces speaks DCE/RPC on TCP port 2012, and that interface runs the Secure Remote Password protocol, SRP-6a, to verify logins.

CVE-2026-123890 is a session key disclosure in that SRP interface. The flaw is in how the rpc_srp_verifier service generates session material: predictable UUIDv1 timestamps in the SRP authentication flow let an attacker work out the shared session key without knowing any password. With the key in hand, the attacker gets an authenticated session as administrator@vsphere.local, the built-in SSO administrator. From there, the directory service's own RPC interface lets them run commands and pull files, including the directory database itself, data.mdb.

Broadcom's advisory is blunt about the blast radius. Every supported vCenter line is affected: 8.0, 9.0, and 9.1, plus the vCenter builds inside VMware Cloud Foundation, vSphere Foundation, and the Telco Cloud products. vCenter 7.0, which left general support in October 2025, is only patched under an extended support contract, and versions 6.5 and 6.7 should be presumed affected. It does not matter whether Enhanced Linked Mode is on or off. The one piece of good news is narrow: ESX itself is not affected, and none of the flaws in this advisory lets code escape from inside a virtual machine onto its host. There is no workaround; Broadcom lists "None" for every CVE, and its FAQ calls the update an emergency change under ITIL.

How it was used in the real world

The group Mandiant tracks as ARMORY CASTLE, formerly UNC5221, did not find vCenter first. They got in through a different zero-day, CVE-2026-102764 in BeyondTrust Remote Support, exploited from late August and patched in October. From the compromised Remote Support appliance they moved laterally to an internal Linux box with no endpoint detection tooling, dropped webshells, escalated to root, and then went looking for VMware. Their scans were specific: ports 9080, 5480, and 2012.

Then they ran BONESAW, a custom exploitation utility written in Go, against the vCenter. BONESAW used CVE-2026-123890 to create a new SSO administrator account, exfiltrate the directory database, and pull out the SAML Security Token Service private signing key. A second utility, SAWSNAKE, turned that stolen key into Golden SAML attacks: forged authentication tokens, MFA bypassed, any user impersonated. The group also enabled SSH on the vCenter appliance, pulled the master symmetric key and the PostgreSQL database credentials, added an SSH key to the root account of an ESXi host, and extracted files from guest virtual machine disks, including Windows Active Directory database files.

Two details in Mandiant's write-up deserve attention. The vCenter they hit was already patched against CVE-2026-59309, a different critical authentication bypass in the same directory service disclosed on July 29. Patching the first flaw did not cover the second. And the BONESAW binary carries a build timestamp of July 13, sixteen days before CVE-2026-59309 was even disclosed. The actor knew about this flaw before the previous one went public. This is the second critical authentication flaw in vCenter's directory service this year, and the people exploiting it were ahead of the disclosure cycle both times.

The patch, and the one upgrade trap

Apply the patch for your line, but read the footnote first if you run Cloud Foundation.

Your version Fixed version
vCenter 9.1.x (VCF / vSphere Foundation) 9.1.1.0200
vCenter 9.0.x (VCF / vSphere Foundation) 9.0.2.0200
vCenter 8.0 8.0 U3m
VMware Cloud Foundation 5.x 5.2.5 (async patching)
vCenter 7.0 Contact Broadcom Support (extended support only)

The updates are cumulative and cannot be installed through vCenter Quick Patch. Virtual machines keep running while vCenter updates, though management access drops briefly and some encryption operations against a Standard Key Provider are unavailable until vCenter is back.

The trap: installing vCenter 8.0 U3m or 9.0.2.0200 blocks a later upgrade to VMware Cloud Foundation 9.1.1. Broadcom's FAQ says to upgrade vCenter to 9.1.1 first and apply the security update there, or wait briefly for VCF 9.1.2. If an upgrade to 9.1.1 is already in flight, check the FAQ before patching in place.

The same advisory fixes four more flaws, all with no workaround: an SQL injection in SDDC Manager (CVE-2026-123891, 8.1), an authenticated command injection in SDDC Manager (CVE-2026-123892, 7.2), a local privilege escalation on the SDDC Manager appliance (CVE-2026-123893, 6.7), and a SAML token signature bypass in VMware Tools for Windows (CVE-2026-123894, 6.7). The same patch window closes all five.

Patching is step one, not the whole job

Here is the part that gets missed when a patch lands and the ticket closes. BONESAW did not just get an admin session. It stole the SAML signing key. A patch fixes the flaw; it does not invalidate a stolen signing key. If your vCenter was reachable during the window from early September to November 19, patching alone does not stop an attacker who already took the key from forging tokens afterwards.

So the job after the patch is a compromise check, not a victory lap. Look for SSO administrator accounts you did not create, SSH enabled on the vCenter appliance when it should be off, new machine accounts in the directory, and SAML assertions that do not line up with real logins. If anything looks wrong, rotate the STS signing certificate and the SSO passwords. Mandiant's VCSA hardening and logging guidance, and the vSphere defender guide, cover what to lock down and what to watch going forward. And while the patch is pending, the exposure is a network question, not a password question: ports 2012, 5480, and 9080 are the ones this actor scanned for, and any of them reachable from where they should not be is a finding on its own.

The durable point

The login page was never the weak point here. Nobody phished a vCenter admin or guessed a password. The directory service answered on a port most people never think about, and the session it handed out was the built-in administrator. vCenter is the asset that manages every other asset, which makes its patch cadence a different kind of commitment than patching a single server. Two critical authentication flaws in the same service in four months, exploited by an actor who was ahead of disclosure both times, is a pattern. Treat the management plane like the keys to the building, because that is what it is, and check the locks after someone has had ten weeks with a copy.

Sources