← cd ../blog

CVE-2026-123890: unauthenticated SSO admin access through vCenter's directory service

An authentication bypass in the VMware Directory Service (CVSS 9.8, CISA KEV) gives unauthenticated attackers SSO admin on vCenter. Root cause, exposure checks, Mandiant's hunting leads and the fix.

Broadcom published VMSA-2026-0009 on 19 November 2026, and CISA added its headline bug, CVE-2026-123890, to the KEV catalog today. It’s an authentication bypass in the VMware Directory Service (vmdir), the component behind vCenter Single Sign-On. Mandiant, who reported it, says a suspected PRC-nexus espionage group was exploiting it as a zero-day from early September. vCenter controls every host and VM in the estate, so this one goes to the top of the queue. And “our vCenter isn’t on the internet” doesn’t settle it: the documented campaign reached vCenter from inside the network.

Scope and ethics. This post covers architecture, exposure checks, detection and remediation for systems you own or are authorized to assess. It contains no exploit code or protocol-level trigger details. Lab validation means confirming builds, firewall rules and detections on infrastructure you control, not running an exploit against anyone’s vCenter.

At a glance

   
CVE CVE-2026-123890 (part of VMSA-2026-0009)
Severity CVSS 3.1 9.8 Critical, AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H (VMware CNA, via NVD)
CWE CWE-288 Authentication Bypass Using an Alternate Path or Channel
Affected vCenter 9.1.x, 9.0.x, 8.0, 7.0; VMware Cloud Foundation 5.x; vSphere Foundation; Telco Cloud Platform / Infrastructure. ESX is not affected
Fixed in vCenter 9.1.1.0200, 9.0.2.0200, 8.0 U3m; VCF 5.2.5; Telco Cloud per KB463116; 7.0 via extended support only
Exploited? Yes. Zero-day use since early September 2026 (Mandiant). CISA KEV, added 20 Nov 2026 (federal due date 23 Nov), forensic triage required
Auth / interaction None / none. Network access to vCenter is enough
Workarounds None (Broadcom)

Description: how the bug works

vCenter’s identity layer is built on vmdir, an LDAP-style directory that stores the vsphere.local SSO domain: users, groups, solution users, and the Security Token Service (STS) signing keys that sign every SAML token vCenter issues. Besides LDAP on 389/636, vmdir exposes a DCE/RPC control interface that Mandiant places on TCP 2012, the “control interface RPC for vCenter Single Sign-On.”

Broadcom’s description is one line: an authentication bypass in the VMware Directory Service that lets “a malicious actor with network access to vCenter” bypass authentication. Mandiant’s write-up adds the architecture. Some RPC operations authenticate with SRP-6a, a password-authenticated key exchange in which both sides derive a shared session key without the password ever crossing the wire. Mandiant calls CVE-2026-123890 an “unauthenticated SRP-6a session key disclosure” in the rpc_srp_verifier service, rooted in a “cryptographic logic and authorization flaw” that involves predictable UUIDv1 timestamp generation. In plain terms: values that should be unpredictable per session aren’t, so an attacker can obtain the session key without knowing any password. The RPC channel then treats the attacker as authenticated. Mandiant says the result is a session “established as administrator@vsphere.local.”

That’s what CWE-288 means here. The normal login path (vSphere Client, STS, MFA if configured) isn’t broken. A second, internal channel into the same directory is, and it skips all of that.

From that session, Mandiant observed the attacker using the vmdir database-copy RPC interface to pull data.mdb, the directory database itself. Whoever has data.mdb has the STS signing material, and with it they can forge SAML tokens for any user (“Golden SAML”), bypassing MFA.

Related bug. Broadcom’s FAQ says this is distinct from CVE-2026-59309 (VMSA-2026-0006, July 2026), an earlier vmdir authentication bypass over LDAP. Mandiant says both bugs share the core trick of obtaining the shared session key without the password, and that the actor knew about CVE-2026-123890 before CVE-2026-59309 was disclosed. Patching for 59309 does not fix 123890.

Impact

  • Full SSO administrator on vCenter without credentials, which means control of every ESXi host and VM it manages.
  • Theft of the directory database and the STS signing keys, which enables forged SAML tokens and MFA bypass until those keys are replaced.
  • Hypervisor and guest compromise. Mandiant documents the actor moving on to ESXi as vpxuser and root, and pulling domain controller secrets (NTDS, SYSTEM, SECURITY hives) directly from guest VM disks. A vCenter compromise can become an Active Directory compromise without touching a domain controller over the network.

Which deployments are exposed

Deployment Status
vCenter 9.1.x below 9.1.1.0200, 9.0.x below 9.0.2.0200, 8.0 below U3m Vulnerable
vCenter 7.0 Vulnerable. End of General Support was 2 Oct 2025; fixes only through extended support contracts
Older, unlisted versions Broadcom’s FAQ says to assume they’re affected
VCF 5.x, vSphere Foundation, Telco Cloud Platform 3.0–5.2, Telco Cloud Infrastructure 3.0 Affected; see the advisory and KB463116
With or without Enhanced Linked Mode Affected either way (Broadcom FAQ)
ESX / ESXi hosts Not affected by this VMSA (but at risk through a compromised vCenter)

Prerequisites: network reachability to the vCenter appliance (Mandiant’s actor specifically scanned for TCP 2012, 5480 and 9080), an unpatched build, no credentials and no user interaction. In the documented intrusion the attacker came in through a different compromised appliance on the internal network, so “not internet-facing” isn’t a reason to wait.

Identifying affected systems

Broadcom’s FAQ suggests checking the build in the vSphere Client (Summary tab of the vCenter object) or with PowerCLI:

# Version and build of each vCenter you manage
Connect-VIServer -Server vcenter.example.internal
$global:DefaultVIServer | Select-Object Name, Version, Build

Compare the result against the fixed versions in the table above, using the vCenter release notes to map version strings to builds. On the appliance shell you can also run vpxd -v.

Then check who can reach the directory service. From a host on a user or server segment that shouldn’t need vCenter management access:

# Benign reachability check against your own vCenter. Any "open" result is a firewall gap.
nmap -Pn -sT -p 22,443,2012,5480,9080 vcenter.example.internal

In a healthy design, only vCenter peers and a small admin/PAW network reach 2012, 5480 and 22. Broadcom’s Ports and Protocols tool lists the legitimate flows for your topology.

Safe validation in an isolated lab

You don’t need to exploit the bug to prove you’re fixed:

  1. Stand up a nested or spare vCenter in an isolated lab on a pre-fix 8.0 build. Run your vulnerability scanner and the PowerCLI check above. Both should flag it. Patch to 8.0 U3m and confirm both stop flagging it.
  2. Apply the appliance-level iptables allowlist Mandiant recommends (inbound DROP by default, allowlisted sources for 443, 2012, 5480 and 22) and re-run the nmap check from a non-allowlisted host. Test this properly before production; Mandiant explicitly asks for “rigorous testing and formal change control.”
  3. Enable SSH through the VAMI and confirm your SIEM alerts on it (see below). Disable it again.
  4. Create and delete a test user in vsphere.local, and confirm the event reaches your SIEM.

Detection and threat hunting

Broadcom’s FAQ has no IOCs. Everything below comes from Mandiant/GTIG’s 20 November write-up on the campaign. Treat it as one actor’s playbook, not a complete list.

  • SSO accounts. New local accounts in vsphere.local, especially in the Administrators group. Mandiant’s actor created an account matching sps-*@vsphere.local. Review Administration > Single Sign On > Users and Groups and compare against your change records.
  • SSH toggling through the VAMI. PUT /rest/appliance/access/ssh enabling SSH, and later disabling it again, is a strong signal on a vCenter where SSH is normally off.
  • Port scans. Internal hosts running SYN scans against 2012, 5480, 9080, 443, 389 and 22. Port 2012 rarely appears outside vCenter-to-vCenter traffic, so scans for it are worth an alert by themselves.
  • Files on the appliance. Mandiant lists webshells (ActiveSer.jsp, u.jsp) and tools staged in /tmp (/tmp/123, /tmp/sss, /tmp/dbr, /tmp/gdb_file), plus /usr/bin/vmtoolscs.
    # quick triage on a vCenter appliance you administer
    ls -la /tmp /var/tmp /dev/shm
    find / -xdev -name '*.jsp' -newer /etc/hostname 2>/dev/null | head -50
    find / -xdev -perm -4000 -type f 2>/dev/null   # Mandiant's actor set SUID on /bin/env
    
  • Unexpected listeners. A backdoored SSH server listening on TCP 33992 or 53. Run ss -tlnp and check for anything you don’t recognize.
  • ESXi. New SSH keys in root’s authorized_keys, SSH logins as vpxuser or root, and hidden directories (Mandiant saw .a) on VMFS datastores.
  • YARA. GTIG published a hunting rule for BONESAW, the exploit tool, keyed on the vmdir database path and the RPC interface UUIDs. Use it from the source post rather than a copy.
  • Logs. vmdir logs usually live under /var/log/vmware/vmdird/ (confirm the path on your build), alongside the VAMI and sshd logs. Mandiant notes the actor may clear bash history, so forward appliance logs to a remote syslog server before an incident, not after.

Preserve evidence before you patch. CISA’s KEV entry requires forensic triage. Snapshot or image the appliance, and collect a support bundle and logs, before you upgrade.

Remediation

  1. Upgrade vCenter to 9.1.1.0200, 9.0.2.0200 or 8.0 U3m; VCF 5.x to 5.2.5; Telco Cloud per KB463116. Broadcom lists no workarounds. Notes from the FAQ:
    • The update isn’t eligible for vCenter Quick Patch. Running VMs keep running, but the vSphere Client is briefly unavailable.
    • 8.0 U3m and 9.0.2.0200 block upgrades to VCF 9.1.1. If you’re mid-upgrade, read FAQ Q39 before choosing a path.
  2. vCenter 7.0: get patches through extended support, or treat migration as urgent.
  3. If vCenter was exposed to untrusted networks before patching, assume the SSO domain is compromised until you’ve hunted. Patching closes the door but doesn’t evict anyone. Broadcom’s FAQ doesn’t cover credential rotation; my recommendation after a confirmed or suspected compromise is to:
    • replace the STS signing certificate, since a stolen data.mdb lets an attacker forge tokens;
    • reset administrator@vsphere.local and every other SSO account;
    • rotate ESXi root passwords and review authorized_keys;
    • treat domain controllers hosted on the cluster as exposed, and plan a krbtgt and privileged-account reset.

Interim hardening (not a fix)

  • Restrict vCenter management ports (443, 2012, 5480, 22) to vCenter peers and a dedicated admin network. Mandiant recommends enforcing this with iptables at the Photon OS level, not through the VAMI, so a compromised VAMI or vSphere admin account can’t turn it off. Their VCSA hardening tool automates it.
  • Block egress from the appliance except to known update, NTP, DNS and identity endpoints.
  • Keep SSH and the shell off on vCenter and ESXi unless a change ticket says otherwise.
  • Broadcom’s FAQ lists perimeter firewalling as a compensating control each organization must assess. It isn’t a substitute for patching, especially since this campaign started inside the network. The same pattern, an exploited edge appliance as the stepping stone into core infrastructure, runs through my NetScaler post and the 2026 RCE field guide.

References

Verify every version against the vendor advisory before acting. Published 20 Nov 2026.