
vCenter zero-day CVE-2026-123890: why patching alone is not enough
Anyone running vCenter must patch now. But anyone who only patches still does not know afterwards whether someone is already inside the system. That is exactly the question I address in this article.
What Broadcom has published
The vulnerability is tracked as CVE-2026-123890 and resides in the VMware Directory Service (vmdir) of vCenter. This is the identity store behind vCenter Single Sign-On. Broadcom rates it 9.8 under CVSS 3.1 and classifies it as an authentication bypass using an alternate path (CWE-288).
Anyone who can reach the system over the network needs no credentials. They gain administrator access to the very console from which every ESX host, every virtual machine and every datastore in the environment can be controlled.
Broadcom published the fix in advisory VMSA-2026-0009, together with four further vulnerabilities in SDDC Manager and VMware Tools for Windows. There is no workaround. In the FAQ for the advisory, Broadcom classifies the vCenter update as an emergency change under ITIL. The advisory itself does not mention active exploitation. That information comes from Mandiant and the Google Threat Intelligence Group.
How the attack unfolded according to Mandiant
Mandiant and Google attribute the attacks to an actor they track as ARMORY CASTLE (formerly UNC5221) and assess as likely China-nexus. This attribution is the assessment of the two companies. I cannot verify it independently from the published data.
Initial access did not go through vCenter itself. The actor first exploited another zero-day vulnerability in BeyondTrust Remote Support (CVE-2026-102764). From there, they moved laterally, obtained root privileges on a Linux appliance in the network and only then attacked vCenter.
Technically, according to Google, this is the disclosure of a session key. A password-based login via the Secure Remote Password protocol (SRP-6a) produces a shared key. vmdir discloses this key without authentication via DCE/RPC. The target is the RPC interface of vCenter Single Sign-On on port 2012. With the key, a session as administrator@vsphere.local can be established without knowing the password.
For this, the actor used a custom tool written in Go. Mandiant calls it BONESAW; the actor itself had named it vcenter-db-teamcheck-go. After the takeover, the actor created SSO administrator accounts, extracted the SSO database (LMDB), took over the SAML Security Token Service, enabled SSH on the appliance and obtained the master key as well as the credentials of the PostgreSQL database.
This sequence is decisive for the assessment. Whoever holds the signing key of the token service can issue valid SAML tokens themselves. This still works after the original vulnerability has been closed.
Affected versions and fixes
Fixed versions per product line:
- vCenter 9.1 (VCF and vSphere Foundation 9.1): 9.1.1.0200
- vCenter 9.0: 9.0.2.0200
- vCenter 8.0: 8.0 U3m (8.0.3.01200, build 25934264)
- VMware Cloud Foundation 5.2: 5.2.5
- Telco Cloud: see Broadcom KB 463116
vCenter 7.0 is missing from Broadcom’s matrix. The FAQ recommends treating this version as affected and requesting a patch through an active extended support contract. Google explicitly describes exploitation against appliances running versions 7.x and 8.x. Older versions such as 6.5 and 6.7 should also be treated as vulnerable. The ESX hosts themselves are not affected by this vulnerability.
There is a catch with the update. vCenter 8.0 U3m and SDDC Manager 9.0.2 block a later upgrade to VCF 9.1.1. Broadcom advises either waiting for VCF 9.1.2 or upgrading vCenter to 9.1.1 first (KB 67077). Anyone planning a migration path should clarify this before the emergency update.
For VMware Tools on Windows (CVE-2026-123894), Broadcom lists versions 13.1.5.0, 12.5.5 and, for 32-bit Windows, 12.4.10. Stopping or uninstalling the tools does not count as a workaround.
Why the July patch does not protect
In July 2026, Broadcom already closed an authentication bypass in the VMware Directory Service with VMSA-2026-0006, CVE-2026-59309. The description sounds almost identical. However, it is a separate vulnerability. Mandiant reports a server that had applied the July patch and was nevertheless taken over with BONESAW.
Anyone who patched in the summer and feels safe as a result is mistaken.
What I would check now
The patch closes the way in. It does not remove an attacker who is already inside, and it does not invalidate stolen keys. Because exploitation began by early September at the latest, according to Mandiant, every vCenter instance that was reachable from a potentially compromised network segment during this period should be examined.
The following points are taken from the guidance published so far on this campaign.
- Review the accounts in vCenter SSO and clarify every administrator account that nobody created intentionally.
- Check the appliance’s VAMI logs for SSH that was enabled and disabled again shortly afterwards.
- Search the network for SSH services on high ports that identify themselves with the banner
SSH-2.0-pulsesecure_23345. - Restrict access to port 2012 to trusted management hosts.
- Check whether BeyondTrust Remote Support is patched against CVE-2026-102764, if the product is in use.
- Bring internal Linux appliances under EDR coverage. According to the report, this is exactly where the actor established itself, because usually nobody looks there.
Before anyone starts searching on a suspicious appliance, volatile data and logs should be secured. A reboot after the update can destroy traces that will later be missing from the investigation.
If a suspicion is confirmed
Then the update is not enough. The signing certificate of the vCenter Security Token Service and all SSO credentials must be renewed. Because, according to Mandiant, the actor also obtained the master key and the database credentials, these must be rotated as well.
There is also the question of what happened between the intrusion and its discovery. Anyone with administrator rights on vCenter could access every virtual machine, take snapshots and copy disks. Whether this happened can only be reconstructed from the environment’s logs. Whether a reporting obligation under the GDPR or the NISG has been triggered depends on the outcome of this reconstruction.
Assessment for SMEs
In many companies, vCenter sits at the centre of IT. If it falls, the attacker does not have one server, but all of them. At the same time, vCenter is an appliance that usually runs no EDR and whose logs are rarely analysed centrally. This combination makes it attractive to espionage actors.
The good news is that the attack requires internal network access. A vCenter does not belong on the internet, and port 2012 does not need to be reachable from every segment. Cleanly separating the management layer reduces the attack surface considerably, also for the next vulnerability of this kind.
The patch is mandatory. It does not answer the question of whether someone has been inside the system in recent weeks. Computerforensic & more GmbH supports the investigation of compromised vSphere environments, from securing the evidence to assessing what an attacker actually achieved.
Sources
- Google Threat Intelligence Group: ARMORY CASTLE Exploits vSphere
- Rapid7: CVE-2026-123890. Critical Authentication Bypass in VMware vCenter
- The Hacker News: Critical VMware vCenter Flaw
- Cyber Kendra: Chinese Hackers Exploited vCenter Zero-Day CVE-2026-123890 Months Before Broadcom Patch
- Rescana: VMware vCenter Auth Bypass Exploited per Google (CVE-2026-123890)



