A compromised vCenter is not just another server: it is the console from which half an infrastructure’s virtual machines are created, moved and deleted. That is why CVE-2026-123890, at 9.8 out of 10, is not a catalogue vulnerability. It is a door that had been open for months.
vCenter is the control plane of vSphere. It holds the inventory of hosts and virtual machines, the templates, the resource policies and the credentials used to manage everything else. An attacker with an administrator session there does not need to break into every server: from the console they can create a machine, mount a disk, read a backup or move a workload.
Broadcom released the patch on 19 November in advisory VMSA-2026-0009. Google Threat Intelligence and Mandiant dated the exploitation to early September: the group Armory Castle, linked to China and formerly tracked as UNC5221, used the flaw as a zero-day for more than two months.
The flaw sits in vCenter’s directory service. It is an SRP-6a session key disclosure over DCE/RPC, on port 2012, that lets an unauthenticated attacker open an authenticated session as [email protected]. From there they can run commands, read the directory database and move files.
Mandiant documented a Go tool called BONESAW that automates the exploitation: it creates an SSO administrator account, extracts the LMDB with the directory data and steals the SAML private signing key from the token service. With that key, another tool, SAWSNAKE, enables Golden SAML attacks: forging tokens, bypassing MFA and impersonating any user.
The impact does not stop at vCenter. The attackers moved to ESXi hosts and copied Active Directory databases from the virtual machines’ disks. They also deployed web shells and backdoors, including one that uses Slack as a command channel. The full chain turns an authentication flaw into control over the virtualisation layer and, from there, over the systems running inside it.
The campaign had two zero-days. The first was CVE-2026-102764, in BeyondTrust Remote Support, used to enter the network. From there the group moved to an internal Linux appliance, escalated to root and prepared the assault on vCenter. The combination explains why the incident went unnoticed: each piece looked like a different problem.
The two-month window is the most uncomfortable detail. A compromised vCenter does not raise an obvious alarm: the session is valid, the SSO account exists and the actions look like those of a legitimate administrator. That is why the review after patching matters as much as the patch.
Broadcom does not confirm the exploitation. Google and Mandiant do. The gap matters for how the incident is treated: the vendor patch is the same, but the forensic review changes if the server was exposed.
The affected versions are vCenter 7.x and 8.0, the 9.0 and 9.1 branches, and the vCenter builds inside VMware Cloud Foundation 5.x and 9.x, vSphere Foundation and the Telco Cloud platforms. The fixed versions are 8.0 U3m, 9.0.2.0200, 9.1.1.0200 and VCF 5.2.5. There is no public patch for vCenter 7.0: only through Broadcom’s extended support.
There is no workaround. Broadcom recommends applying the patch and restricting network access to ports 2012, 443, 5480 and 22. If the server has been exposed, the patch does not erase what already happened: review the SSO accounts, look for names matching the sps- pattern cited by Mandiant and check whether anyone has assigned the Global Administrator role or accumulated more privileges than they need. If there are signs of compromise, the SAML key must be treated as stolen.
The lesson is the usual one in virtualisation: the management plane concentrates more power than any individual server. A vCenter with SSO administrator access is a master key. And this time, that key was on port 2012.