9.8
CVSS 3.1 score Broadcom gives the vCenter flaw
Sep 2026
when exploitation began, per Mandiant forensics
7.0
vCenter branch with no public fixed build
4
inbound ports Mandiant would leave open on vCenter
The patch closes the door, the key already left
Patching vCenter for CVE-2026-123890 stops the next intrusion. It does nothing about the last one, because what the attackers took from vCenter was its identity, and identity does not expire when the build number changes.
That is the part of this week's VMware news we want every infrastructure team to sit with. The flaw is an authentication bypass in the VMware Directory Service, the component that backs vCenter single sign on. In the intrusions Mandiant and Google Threat Intelligence describe, the attackers used it to create an SSO administrator, copy the directory database and extract the key vCenter uses to sign its login tokens. With that key they forged administrator logins at will. Then they walked down into the hypervisors and copied Active Directory files out of the virtual machines.
Every one of those steps leaves something behind that a patch does not touch. The administrator account is still there until someone deletes it. The directory copy is on someone else's disk for good. The signing key keeps signing for whoever holds it until it is replaced, and replacing it is a separate job that nobody runs by accident.
So this week has two questions. Is every vCenter we run on a fixed build? That one has a clean answer by tonight. And was any vCenter of ours reachable, since early September, from a system we would not stake the estate on? If the answer to the second is yes, or we do not know, the job is not a patch. It is rebuilding trust in the management plane.
What Broadcom and Mandiant say about CVE-2026-123890
Broadcom published VMSA-2026-0009 on 19 Nov 2026. It covers five issues: CVE-2026-123890 in vCenter, three flaws in SDDC Manager, and a SAML token signature bypass in VMware Tools for Windows. The advisory's question and answer page says the vCenter flaw allows an attacker with network access to vCenter to bypass authentication and gain unauthorized access. It lives in the VMware Directory Service. Broadcom lists no workaround, says ESX itself is not affected, and says the issue is present whether or not Enhanced Linked Mode is in use. It also calls the vCenter fix an emergency change under ITIL, which is about as loud as a vendor's change language gets.
Rapid7's analysis gives the score: 9.8 under CVSS 3.1, classed as CWE-288, authentication bypass through an alternate path or channel, exploitable without credentials. It lists the fixed builds. vCenter 8.0 is fixed in 8.0 U3m, which is also the update for VCF 5.x, fixed as a release in 5.2.5. The 9.0 line is fixed in 9.0.2.0200 and the 9.1 line in 9.1.1.0200. vCenter 7.0 has no public fixed build: Broadcom points customers with extended support to its support process, since vSphere 7 left general support on 2 Oct 2025. Broadcom says versions 6.5 and 6.7 should be presumed affected.
-
Component //
VMware Directory Service in vCenter -
Needs //
network access to vCenter, no credentials -
Fixed in //
8.0 U3m, 9.0.2.0200, 9.1.1.0200; VCF 5.2.5 -
vCenter 7.0 //
no public fixed build, extended support only -
CVSS //
9.8 (3.1, Broadcom) -
Exploited //
Yes, per Mandiant, from early September 2026
Two operational details will bite teams that move fast. The vCenter updates are not eligible for vCenter Quick Patch. And Broadcom warns that 8.0 U3m and 9.0.2.0200 create what it calls a Back-in-Time problem: they block an upgrade to VCF 9.1.1, so anyone mid way through a 9.1.1 upgrade should either upgrade vCenter to 9.1.1 first and patch there, or wait briefly for VCF 9.1.2. Broadcom is clear that this is a new and distinct issue from CVE-2026-59309, the earlier Directory Service bypass it fixed this year. A vCenter that took the earlier fix still needs this one.
On exploitation, the sources say different things, and the difference matters. BleepingComputer reports that Broadcom has not confirmed the flaw was exploited. Rapid7 notes Broadcom credits Mandiant with reporting it privately. And Google Threat Intelligence's write-up, published the same day, says forensic evidence shows ARMORY CASTLE exploited the vulnerability in early September 2026 using a custom tool. GTIG attributes the activity to ARMORY CASTLE, a suspected PRC-nexus espionage actor formerly tracked as UNC5221. We repeat that attribution as GTIG states it, with its hedge.
GTIG ties this campaign to its 15 Oct 2026 report on a zero-day in BeyondTrust Remote Support, CVE-2026-102764, which BeyondTrust patched in BT26-05. That is how the attackers got in. The vCenter bug is how they got everything else. Readers who followed our Ivanti Connect Secure write-up will recognise the shape: an edge appliance first, then the systems behind it that nobody watches as closely.
How ARMORY CASTLE went from an appliance to Active Directory
GTIG's account reads like a tour of the management plane. We have kept it at the level of what each stage achieved, and left the tool internals in the original report.
Figure 01
From a support appliance to the domain
- 01
In //
A BeyondTrust zero-day opens the perimeter.
- 02
Hide //
Root on internal Linux boxes with no EDR.
- 03
Take //
SSO admin, directory copy, signing key.
- 04
Forge //
Golden SAML logins as administrator.
- 05
Copy //
AD database files out of guest VMs.
Result //
The hypervisor layer handed over the domain.
After the BeyondTrust appliance, the attackers placed web shells on an internal Linux appliance, escalated to root and scanned for VMware systems. GTIG notes that they favoured internal Linux appliances without endpoint detection. That choice is deliberate and it is common: appliances are the part of the estate where security tooling is not allowed to run, and the part teams are least likely to log in to by hand and notice something odd.
At vCenter, a custom Go utility GTIG calls BONESAW exploited CVE-2026-123890. GTIG says it created an SSO administrator account, exfiltrated the directory database and extracted the token signing key of the vCenter Security Token Service. From there the attackers enabled SSH on the appliance through its management API, created a local administrator, and took the vCenter master key and its database credentials.
Then they stopped needing the bug. A second tool, SAWSNAKE, forged administrator SAML assertions with the stolen signing key and turned them into administrator sessions: a Golden SAML attack against vCenter. From vCenter they authenticated to the ESXi host client as vpxuser, the account vCenter itself uses to manage hosts, and connected to the host over SSH as root. Using what GTIG calls a suspected VMware specific toolset, they browsed the datastores and pulled files directly out of guest virtual machines. The list GTIG gives includes the Windows NTDS database and the SYSTEM and SECURITY hives, which is to say the material needed to recover password hashes for the whole Active Directory domain.
Nothing in that last stage needed a Windows exploit, a login to a domain controller, or a single alert from the endpoint agent on that controller. The attackers read the disk from underneath. For persistence GTIG lists a backdoored SSH server listening on ports the attackers chose, and a backdoor that takes its commands through Slack and is restarted by a systemd service every 50 minutes. GTIG does not give a victim count, a dwell time or target sectors, and we do not guess at them.
Why a stolen signing key outlives the patch
The signing key is the whole story, so it is worth being precise about what it is. vCenter's Security Token Service issues the SAML tokens that prove who a user is to vCenter and the services around it. Whoever holds the private key can mint a token that says anything: any user, any group, administrator included. vCenter checks the signature, finds it valid, and lets the bearer in. No password is checked, no MFA prompt is shown, and the login looks like an ordinary one.
The patch for CVE-2026-123890 closes the path that let BONESAW reach the directory without credentials. It does not change the key. If the key left the building in September, a forged token minted in December is just as valid as one minted the day it was stolen. The same goes for the directory database: it holds the SSO users, groups and credential material, and a copy taken before the patch does not become harmless after it.
We have seen this pattern before in identity products. A flaw lets an attacker steal or bypass the thing that signs trust, the vendor fixes the flaw, and the organisations that only patched are still exposed to whoever kept the material. We made the same argument for SAML on the edge in our SonicWall SSL VPN write-up, and for the token side of the problem the stolen token response guide covers why a password reset does not evict the holder.
GTIG's post does not tell defenders to rotate the STS signing certificate or to reset SSO credentials. Its advice is to patch, firewall the appliance, filter egress and forward logs. That advice is right for the next attacker. For an organisation that cannot rule out the last one, our recommendation goes further: replace the STS signing certificate, review and remove SSO and local administrator accounts that nobody can account for, rotate the credentials vCenter holds for hosts and integrations, and treat the domain as exposed if any vCenter could reach the disks of domain controllers. That is a larger change window than a patch. It is also the only step that takes the stolen key out of play.
The ESXi detail sharpens the point. Broadcom says ESX is not affected by this advisory, and that is true of the code. It is not true of the outcome. Once vCenter is owned, its own management account on every host is the attacker's account too. We made a related argument about virtualisation management in our Xen Orchestra write-up: whoever runs the hypervisor console owns every machine on it, whatever those machines' own defences say.
How to tell if your vCenter was touched
Start with what GTIG published. It released YARA rules for the six named tools, a table of network indicators and a table of file indicators, with more in a collection for registered users. Run the YARA rules against the vCenter appliance and any Linux appliances that could reach it. Then look at behaviour, because a careful operator will not reuse the same files twice.
Four signals from GTIG's report are worth hunting for directly. SSH being enabled on the vCenter appliance through its management API, which GTIG names as a hunting lead. SSO administrator accounts and local appliance accounts created since September that no change ticket explains. Outbound connections from vCenter or ESXi to services a hypervisor has no business talking to, Slack among them. And SSH or other listeners on ports nobody configured, which on a hypervisor appliance should be a very short list.
The network signal is the one most teams can check today. GTIG says the attackers scanned for VMware management ports including 2012, 5480 and 9080, and that port 2012 is the vCenter SSO interface BONESAW used. On a healthy estate, the only systems that talk to vCenter on that port are other vCenter and SSO components. The rule below is a starting point for flagging anything else.
Figure 02
vCenter SSO port reached from outside the allowlist
Sigma
title: Connection to vCenter SSO interface from an unexpected source
status: experimental
description: Flags connections to TCP 2012 on vCenter from hosts that are not other vCenter or SSO nodes.
logsource:
category: firewall
detection:
selection:
dst_ip:
- 192.0.2.10
dst_port: 2012
filter_known:
src_ip|cidr:
- 192.0.2.0/28
condition: selection and not filter_known
falsepositives:
- Other vCenter or SSO nodes missing from the allowlist
level: high Many teams will find the honest answer is that they have no logs to run it on. GTIG recommends remote syslog forwarding of shell command execution and dropped packets from the appliance, and iptables connection logging into a SIEM. If none of that was in place in September, the absence of an alert proves nothing, and the decision about rotating the signing key should be made on reachability, not on evidence of absence.
How we test the management plane on an engagement
On internal network tests and assumed breach engagements, the management plane is one of the first things we map. The question is not whether vCenter is patched. It is which networks can reach it, and on which ports. On a sanctioned engagement we start from a user segment and from the kind of appliance an attacker lands on first, and we record every route to the vCenter web interface, the appliance management interface, the SSO port and SSH. The internal network first hour guide covers the mapping, and the remote admin service checks cover how we read SSH exposure without logging in.
What we report is usually a flat network problem, not a VMware problem. Hypervisor management sits on a network that backup servers, monitoring, jump hosts and a few appliances can all reach, and those appliances run software the customer cannot install an agent on. That is exactly the path GTIG describes. The fix is a management network with a short, written allowlist, which happens to be what GTIG recommends for the appliance itself: default deny, with inbound TCP 443, 2012, 22 and 5480 allowed only from named sources, plus egress filtering.
We also test what happens after vCenter. On engagements where the client allows it, we show the reach from a vCenter administrator to the domain controller's disk, without copying any data, because it is the finding that changes how a board thinks about hypervisor access. A domain controller that runs as a virtual machine is only as protected as the people and systems that can manage the host it runs on.
For the outside view, our open source recon tool, atarus-recon, maps what an organisation exposes on the internet. A vCenter or appliance management interface on that list is a finding on its own, whatever its build number.
What is still coming
Three things are likely in the next few weeks. More detail on scope: GTIG gives no victim count, and organisations that ran the affected BeyondTrust appliance will now be checking their vCenter logs with a specific question in mind. Wider use: the exploitation so far is one espionage actor with custom tooling, but a public patch tends to shorten the time to a public reproduction, and Directory Service bypasses have now appeared twice in one year. And government guidance: we expect the flaw to be considered for the CISA Known Exploited Vulnerabilities catalog, and a listing would set a deadline for federal agencies.
The longer trend is the one GTIG keeps documenting. This actor goes to appliances first and to hypervisors second, because both sit outside the reach of endpoint detection and both hold the keys to systems that do have it. Teams that measure their coverage by agent install counts on Windows servers will keep missing the layer underneath them.
What to do today
-
Patch every vCenter to 8.0 U3m, 9.0.2.0200 or 9.1.1.0200. VCF 5.x takes 8.0 U3m. Quick Patch does not apply, and anyone mid upgrade to VCF 9.1.1 should read Broadcom's Back-in-Time note first. vCenter 7.0 needs the extended support route.
-
Cut network access to vCenter down to a written allowlist. Inbound 443, 2012, 22 and 5480 from named admin and vCenter hosts only, as GTIG recommends, plus egress filtering on the appliance.
-
Run GTIG's YARA rules and indicators against vCenter and the Linux appliances that can reach it. Then hunt for SSH enabled on the appliance, unexplained SSO or local admin accounts, and outbound traffic to services like Slack.
-
If any vCenter was reachable from a system you do not fully trust, replace the STS signing certificate. Remove unexplained accounts and rotate the credentials vCenter holds. The patch does not do this for you.
-
If domain controllers run as virtual machines on a host that vCenter managed, treat the domain as exposed. Plan the Active Directory credential rotation, including the accounts that sign Kerberos tickets, as part of the same response.
-
Start forwarding shell and connection logs from vCenter and ESXi to your SIEM. The Sigma rule above needs data to run on.
SOURCES
- Broadcom VMSA-2026-0009 Q&A // CVE-2026-123890 is a vCenter authentication bypass needing only network access, no workaround, ESX not affected
- Google Threat Intelligence // ARMORY CASTLE exploited the flaw from early September 2026, the intrusion chain, tooling and hardening advice
- Rapid7 // CVSS 9.8 under 3.1, fixed builds per branch, vCenter 7.0 through extended support only
- BleepingComputer // independent report of the patch, the attribution and that Broadcom has not itself confirmed exploitation