Active Directory & Identity

Patching vCenter Does Not Take Back the SAML Signing Key ARMORY CASTLE Already Copied

Dark cyberpunk illustration of a heavy iron key resting on a black pedestal in front of a sealed vault door, with a glowing amber duplicate of the same key streaming away into the dark over cyan circuit traces.

A copy of data.mdb left a vCenter appliance in early September, and that one file is the reason patching this week is only the first of two jobs. data.mdb is the VMware Directory Service database. The SAML signing key for the whole single sign-on domain lives inside it. An attacker holding that key can mint an administrator assertion for vCenter at will, with no password, no multi-factor prompt, and no further need for the vulnerability that handed them the file.

Broadcom shipped the fix on November 19, 2026 as VMSA-2026-0009. CISA added CVE-2026-123890 to the Known Exploited Vulnerabilities catalog the next day with a federal due date of November 23, which is three days. The catalog entry carries a field most operators never read. In catalog version 2026.11.20 this CVE is marked "forensicTriage": "Yes", as are three of the four entries added that day. Under BOD 26-04, that flag moves the required action from install the update to install the update and then go looking.

Google Threat Intelligence Group and Mandiant attribute the zero-day exploitation to ARMORY CASTLE, a suspected China-nexus espionage group previously tracked as UNC5221. Their write-up is the technical basis for everything below, and its most uncomfortable detail is the target. The vCenter they reached in early September had already been patched for CVE-2026-59309, the vmdir authentication bypass disclosed in July.

Who has to act on this, and who does not

The flaw is in vCenter Server 7.x and 8.x appliances and in the vCenter component of VMware Cloud Foundation and vSphere Foundation 9.0 and 9.1. If you run no vCenter, this one is not yours. Per Broadcom's advisory question-and-answer page, ESX itself, NSX, VCF Operations, VCF Automation, Workstation and Fusion are not affected by this CVE. A standalone ESXi host with no vCenter in front of it has nothing to patch here.

Two groups should read the rest carefully. The first runs vCenter 7.0, which reached end of general support on October 2, 2025 and has no listed fixed version. Extended-support customers are told to contact Broadcom. Everyone else on 7.0 has an actively exploited unauthenticated administrator bypass and no patch, and Broadcom says to presume 6.5 and 6.7 affected, because versions past end of general support are not evaluated. In the small and mid-sized environments I assess through Red Hound, vCenter 7.0 is still the common case, because the upgrade is a project and the hypervisor has been reliable for years.

The second group patched in July for CVE-2026-59309 and recorded the hypervisor as handled. The July bug was an LDAP flaw in the vmdir daemon on TCP 389 and 636, reached with malformed authorization tokens. This one is a separate code path, and Broadcom confirms the earlier fix does not cover it.

What the bug does, and the port it does it on

Mandiant describes CVE-2026-123890 as an unauthenticated Secure Remote Password session key disclosure over DCE/RPC. The vulnerable interface is rpc_srp_verifier, served by vmdird on TCP 2012. SRP lets a server prove knowledge of a password verifier without the password crossing the wire, and both sides derive a shared session key from the exchange. The implementation generated part of that exchange from UUID version 1 values, which embed a timestamp and are therefore predictable. Predict those values and you derive the shared key with no password at all.

What comes out of it is a fully authenticated session as administrator@vsphere.local. From there the same RPC surface serves file retrieval, so the attacker can pull /storage/db/vmware-vmdir/data.mdb off the appliance through the vmdirdbcp interface. NVD scores it CVSS v3.1 9.8, vector AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H, under CWE-288.

Port 2012 is the operational point. Hardening guides and firewall change tickets talk about 443, because that is where the vSphere Client lives. If your management VLAN access-control list permits any-to-vCenter because it is only the admins' subnet, 2012 is open to everything in that subnet, including the Linux appliance with no endpoint agent that ARMORY CASTLE used as a launch point.

Patch sequence, and the two traps in it

The fixed releases are vCenter Server 8.0 U3m, VCF and vSphere Foundation 9.0.2.0200, and 9.1.1.0200. VCF 5.x customers apply vCenter 8.0 U3m through the asynchronous patching process in KB 88287 and move SDDC Manager to 5.2.5. Telco Cloud Platform and Telco Cloud Infrastructure take 8.0 U3m via KB463116. Broadcom lists no workaround and classes the change as an emergency.

Trap one: these patches are not eligible for vCenter Quick Patch, so budget for the full update path and a management-plane outage. Running virtual machines and containers stay up; your ability to manage them does not.

Trap two is the one that will bite anyone mid-migration. vCenter 8.0 U3m and 9.0.2.0200 both add a Back-in-Time restriction that blocks a later upgrade to VCF 9.1.1. Broadcom's guidance is to wait for VCF 9.1.2, or to finish the move to 9.1.1 first and then apply the security update. Decide which of those you are doing before you start, because the security patch is the step that removes the other option.

The same bundle carries CVE-2026-123894, a SAML signature bypass in VMware Tools for Windows, fixed in 13.1.5.0 and 12.5.5. Broadcom is explicit that disabling guest operations is no workaround and that guest aliases already created survive the update. Run VGAuthCLI.exe list inside Windows guests and read what is there.

The three accounts and one key the patch leaves behind

This is the part the headline coverage skips, and it is why CISA set the forensic-triage flag. Mandiant documents a Go utility called BONESAW whose csa subcommand does four distinct things with one exploit: it creates a single sign-on administrator account, exfiltrates the directory database, extracts the Security Token Service SAML signing key from it, and creates a backdoor machine account in vmdir. A companion tool, SAWSNAKE, is a Go port of Horizon3's public vcenter_saml_login.py and turns that extracted key into forged administrator assertions, which bypasses multi-factor authentication because the assertion is already signed.

Four artifacts, and the update removes none of them:

  • An SSO administrator account created through the exploit, with any name the actor chose.
  • A local administrator named with an sps- prefix at @vsphere.local, which the actor used to log in after enabling SSH through the appliance management API with PUT /rest/appliance/access/ssh.
  • A backdoor machine account in vmdir, which appears in the service-principal list rather than in the user list the vSphere Client shows you.
  • The STS signing certificate and private key, which are now outside your estate.

Start with the accounts, because the enumeration takes minutes:

# On the VCSA appliance shell, as root.
# Members of the SSO Administrators group.
/usr/lib/vmware-vmafd/bin/dir-cli group list \
  --name Administrators --login administrator@vsphere.local

# Service principals. A backdoor machine account created through
# vmdir shows up here, not in the vSphere Client user list.
/usr/lib/vmware-vmafd/bin/dir-cli service list \
  --login administrator@vsphere.local

Then the role grants, because BONESAW's documented final step was Global Administrator on the vCenter root folder:

Connect-VIServer -Server vcsa.internal.example
$global:DefaultVIServer | Select-Object Name, Version, Build

# Every principal holding a role on the root folder, which is
# where the propagating Global Administrator grant was made.
Get-VIPermission -Entity (Get-Folder -NoRecursion) |
  Select-Object Principal, Role, Propagate, IsGroup |
  Sort-Object Principal

The signing key is the harder item and the one with no checkbox. Nothing on the appliance tells you whether it was copied, and Mandiant's report does not tell you to replace it. Replace it anyway if you find any of the account artifacts above, or if TCP 2012 was reachable from a segment you cannot fully account for between early September and the day you patched. A forged assertion signed with a valid STS key looks like a correct login in every log you have. Rotating the certificate is a disruptive but well-documented vCenter procedure, and a maintenance weekend beats discovering in February that the key was still live.

Hunt the appliances, then the hosts

The intrusion Mandiant traced did not stop at vCenter. The actor authenticated to ESXi as vpxuser, added a public key to root's authorized_keys, staged stolen files in a hidden .a directory on a VMFS volume, and used a VMFS reader to pull NTDS, SYSTEM, SECURITY and DNS data straight out of guest virtual machines without ever logging into them. Persistence landed as LAZYBONES at /usr/bin/vmtoolscs, a Go backdoor with Slack command and control, restarted by a systemd unit every 50 minutes.

# On the VCSA and on every Linux appliance in the same segment.
for p in /tmp/123 /tmp/123.rar /tmp/123.key /tmp/sss /tmp/dbr \
         /tmp/rar /tmp/gdb_file /usr/bin/vmtoolscs; do
  [ -e "$p" ] && echo "PRESENT $p"
done
find / -xdev -perm -4000 -name env 2>/dev/null   # gdb set SUID on /bin/env
ss -lntp 2>/dev/null | grep -E ':(53|33992)\b'   # BEARBONES listeners

# On every ESXi host.
for v in /vmfs/volumes/*/; do
  [ -d "$v.a" ] && echo "STAGING DIR $v.a"
done
cat /etc/ssh/keys-root/authorized_keys 2>/dev/null

Two caveats. Those filenames come from one intrusion and are cheap to change, so an empty result is weak evidence rather than a clean bill. The actor also paired this campaign with CVE-2026-41709, an ESX logging flaw used to thin the audit trail, so silence in ESXi logs counts for less than usual.

Put a real firewall in front of TCP 2012

Mandiant published a VCSA hardening and logging tool alongside the report, worth reading even if you never run the script. The appliance management interface firewall cannot express per-port rules, and an attacker who reaches that interface can switch it off. Enforcement belongs in iptables on the appliance, where vSphere administrator credentials cannot reach it: default-deny inbound on 443, 2012, 22 and 5480 with an allowlist of your jump hosts, a default-deny egress policy so a compromised appliance cannot reach Slack or copy archives out, and remote syslog so the shell commands and dropped packets land somewhere the appliance owner cannot edit. Test it in a change window; a bad rule on vCenter is its own outage.

For a ten-person IT team the cheap version of all three is a single access-control list that permits 443, 2012, 22 and 5480 only from one administrative jump host, plus a syslog destination. That is an afternoon of work and it would have broken the chain Mandiant describes at the second step.

Count the accounts in vsphere.local before you call it patched

Install VMSA-2026-0009 this weekend if you run 8.x or 9.x, and open a support case today if you run 7.0, because there is nothing else coming for it. Then do the second job: list the SSO Administrators group, list the service principals, read the role grants on the root folder, and compare all three against what you expect to be there. If anything is unaccounted for, treat the STS signing key as copied and plan its replacement. One intrusion is documented in public, Broadcom's advisory does not discuss exploitation, and nobody has said how many organizations were hit. That is why the count has to come from your appliance rather than from a news cycle.

If you want a second set of eyes on that count, or an incident response plan written before the next management-plane zero-day, that is work we do.

Need a compromise assessment after an emergency patch?

We help organizations work out whether a patched zero-day left anything behind, and build the incident response playbook for the next one. Book a session with our team.