Automated Threat Modeling

Brief Summary: VMware vCenter CVE-2026-123890, a Directory Service Authentication Bypass Exploited by ARMORY CASTLE

A brief summary of CVE-2026-123890, a CVSS 9.8 authentication bypass in the VMware Directory Service of vCenter that Google reports was exploited by ARMORY CASTLE. Covers the reported root cause, affected versions, and the detection guidance published by Mandiant and GTIG.

CVE Analysis

11 min read

Gaurav Sarraf
Gaurav Sarraf

2026-11-19

Brief Summary: VMware vCenter CVE-2026-123890, a Directory Service Authentication Bypass Exploited by ARMORY CASTLE

Introduction

An attacker who can reach a vulnerable vCenter over the network can obtain an [email protected] session without a password, and Google Threat Intelligence Group (GTIG) reports that the threat group ARMORY CASTLE has already done this in real intrusions. Because vCenter is the control plane for ESXi hosts and the guest workloads on them, the intrusion Mandiant documented went from this bypass to root on ESXi and the theft of Active Directory database files straight off guest disks.

Technical information

Classification and scoring

VMware describes CVE-2026-123890 as an authentication bypass in the VMware Directory Service of vCenter, classified as CWE-288 (Authentication Bypass Using an Alternate Path or Channel). The vendor advisory states that a malicious actor with network access may bypass authentication and obtain unauthorized access to the system.

VMware assigned a CVSS 3.1 score of 9.8 Critical with the vector AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H: network attack vector, low complexity, no privileges, no user interaction, unchanged scope, and high impact to confidentiality, integrity and availability. NVD published the entry on November 19, 2026, with VMware as the source. NIST had not yet supplied its own assessment. The CVE data lists the vendor as "Unknown", but the advisory is published by Broadcom for VMware products.

Root cause

According to the Mandiant and GTIG material, the bug is an unauthenticated SRP (version 6a) session key disclosure over DCE/RPC. It is reached through two RPC interfaces exposed by vmdird, the directory service daemon:

  • rpc_srp_verifier (interface UUID c7e94609-6ab0-4767-8e92-b6485bff1649)
  • vmdirdbcp (interface UUID 72c780fe-af6b-42ba-9b02-f61f4ccc98ed)

Both are served on TCP 2012, the vCenter SSO control RPC port. Leaking the session key lets the attacker skip the password proof that SRP is meant to enforce and end up with an [email protected] session without knowing the password. That matches the CWE-288 "alternate path" classification. The UUIDs above come from GTIG's BONESAW YARA rule, which also embeds the SRP prime N and the vmdir database path.

The public sources do not document the exchange at the message level, and we are not going to reconstruct it here. What is established is the transport (DCE/RPC on TCP 2012), the interfaces involved, and the outcome (a passwordless SSO administrator session).

Attack flow observed in the wild

Mandiant places exploitation in or after early September 2026. The sequence below is the intrusion GTIG and Mandiant reported, not a generic model.

  1. Foothold on an internal Linux appliance. The exploit was launched from a compromised internal Linux appliance that had no EDR, not from vCenter itself. The actor gained root there by writing a base64-decoded gdb script and attaching gdb to a process to set the SUID bit on /bin/env:
echo IWNobW9kIHUrcyAvYmluL2VudgpxCg==|base64 -d > /tmp/gdb_file

Tools were then pulled into /tmp with curl -k.

  1. Reconnaissance. The actor ran nmap SYN scans for VMware-specific ports and requested URIs such as /sdk and /ui. Commands recovered by Mandiant:
/usr/sbin/nmap -sS -p 443,2012,389 --open <REDACTED>.0/24 /usr/sbin/nmap -sS -p 443,2012,5480 --open <REDACTED>.0/24 /usr/sbin/nmap -sS -p 9080,2012 --open <REDACTED>.0/24
  1. Exploitation with BONESAW. BONESAW is a Go tool (staged as /tmp/123, with an SSH private key at /tmp/123.key) that attacks vCenter remotely over TCP 2012. Its build path is E:/CodeFiles/VulnScan/Vcenter/go_teamcheck/ and its internal name is vcenter-db-teamcheck-go. Its main csa subcommand creates an SSO admin account, with a default bridge identity of [email protected] and a default domain of vsphere.local. In the observed case the new account was named like sps-<redacted>@vsphere.local. BONESAW can also create a backdoor machine account through vmdir, and grants Global Administrator on the vCenter root folder through the vSphere SOAP API.

  2. Credential and key extraction. BONESAW writes output under <output_root>/temp/extracted_tenant_signer_<host>/, including anon_<host>_data.mdb (a copy of /storage/db/vmware-vmdir/data.mdb) and tenant credential files with .key.pem, .key.der, .leaf.pem, .chain.pem and .context.json extensions.

  3. Golden SAML. The companion tool SAWSNAKE (/tmp/sss) forges SAML 2.0 assertions with the stolen STS signing key and requests sessions through https://<vcenter>/ui/saml/websso/sso. It writes idp_key.pem, trusted_cert_2.pem and domain.txt, which the actor later deleted with rm.

  4. Appliance shell access. The actor sent a PUT to /rest/appliance/access/ssh to enable SSH through VAMI, logged in with the new account, stole the vCenter master symmetric key and the PostgreSQL password file, and then disabled SSH again.

  5. Movement to ESXi. The actor logged into the ESXi Host Client as vpxuser, used root SSH and WinSCP, added a key to root's authorized_keys, created a hidden .a directory on a VMFS volume, extracted vmfs-pydist.tar.gz, and used vmfs_browser.py to pull NTDS, SYSTEM and SECURITY hives from guest disks into o.tar.gz.

  6. Persistence and tooling. Additional implants included BEARBONES (a backdoored Dropbear SSH server at /tmp/dbr), LAZYBONES (/usr/bin/vmtoolscs, persisting through a systemd service with Slack as C2), and the webshells SLIMSAVAGE (ActiveSer.jsp) and CAGEMATCH (u.jsp).

Detection methods

Where detection stands

At disclosure (November 19, 2026), Sigma Watch reports that no public Sigma, Elastic, Splunk ESCU, YARA, Microsoft Sentinel, Snort or Suricata rule references CVE-2026-123890. There is no signature labelled with this CVE that can simply be switched on. The usable material comes from Mandiant and GTIG, and it falls into four layers: finding vulnerable vCenters, watching the network path the exploit needs, auditing the vCenter control plane for what happens after the bypass, and hunting attacker tooling on the staging host.

1. Find exposed vCenters

Broadcom's VMSA-2026-0009 Q&A explains how to read the vCenter build. It appears on the vSphere Client Summary tab. In PowerCLI, after Connect-VIServer, the build is in $global:DefaultVIServer.Build and the version in $global:DefaultVIServer.Version. Compare these with the fixed builds listed below. Any vCenter below them is a candidate for hunting, especially if it was reachable from internal Linux appliances in or after early September 2026.

2. Network layer: TCP 2012 is the choke point

Very few hosts need to talk to TCP 2012 in normal operation, which makes it a strong signal.

  • Unexpected peers on 2012. Alert on any connection to vCenter TCP 2012 from a source that is not a known management host or vCenter peer.
  • Reconnaissance. A single internal host sweeping a subnet for 2012 together with 5480, 9080 or 443 (see the nmap commands above) is unusual and worth flagging, as are requests for /sdk and /ui from unexpected sources.
  • RPC interface UUIDs. Any rewritten exploit would still need to bind to the vmdirdbcp and SRP verifier interfaces, so their UUIDs are good pivots if your network tooling decodes DCE/RPC bind requests on port 2012. This is an inference from GTIG's YARA rule content, not a published network signature.
  • Turn the appliance into a sensor. The native VAMI firewall does not log denied connections remotely. Mandiant recommends its open-source VCSA Hardening and Logging Tool (github.com/mandiant/vcsa-hardening-tool). It applies default-deny iptables rules on Photon OS, including a drop on 2012 for sources not on the allowlist, and an iptables-log-bridge.sh service forwards blocked attempts to your SIEM. Mandiant notes this would have turned the recon scans into early-warning alerts. It also forwards shell command execution to syslog, so evidence survives deletion of .bash_history. Test it under change control first.
  • BEARBONES SSH backdoor. The backdoored Dropbear server announces itself with the banner SSH-2.0-pulsesecure_23345 and was started with /tmp/dbr -R -p 53 and /tmp/dbr -R -p 33992. Look for SSH banners on nonstandard ports and SSH answering on TCP 53. GTIG found a suspected obfuscation network of compromised firewalls and routers by searching for this banner on high ports.
  • Network IoCs. GTIG lists these suspected obfuscation network IPs: 103.147.251.5, 103.190.38.3, 118.174.128.34, 118.174.128.36, 118.174.128.38, 176.9.89.66, 185.71.15.109, 220.121.128.41, 222.108.77.44, 61.216.56.127, 81.45.89.43. A fuller set is in GTIG's GTI Collection on VirusTotal (registration required).
  • Slack traffic from appliances. The LAZYBONES YARA rule keys on Slack API endpoints such as conversations.history, so unexpected Slack API traffic from an appliance is another lead.

3. vCenter control plane audit

Changes made after the bypass are often the most reliable evidence.

  • New SSO administrators. Review all SSO and vmdir accounts, including machine accounts, for entries like sps-<redacted>@vsphere.local. Look for unexpected Global Administrator permissions on the vCenter root folder.
  • SSH toggled through VAMI. Look for a PUT to /rest/appliance/access/ssh enabling SSH followed shortly by disabling it. SecurityOnline's hunting checklist calls out this on/off pattern in VAMI logs, and Google SecOps detects it as "VMware vCenter VAMI API SSH Enablement".
  • Golden SAML. Administrator sessions obtained through /ui/saml/websso/sso with no matching legitimate identity provider login are a red flag.
  • ESXi changes. Watch for changes to root's authorized_keys, dot directories under /vmfs/volumes/, logins as vpxuser to the ESXi Host Client, and the names vmfs-pydist.tar.gz, vmfs_browser.py and o.tar.gz.

4. Host hunting on the staging appliance

BONESAW runs from a separate Linux host, so it will not be found on vCenter. Hunt on internal Linux appliances, especially those without EDR, for ActiveSer.jsp, u.jsp, /tmp/123, /tmp/123.key, /tmp/sss, /tmp/dbr and /usr/bin/vmtoolscs. Search live disk and deleted file or filesystem timeline artifacts for the BONESAW output tree (extracted_tenant_signer_<host>, anon_<host>_data.mdb, tenantcredential files) and SAWSNAKE outputs (idp_key.pem, trusted_cert_2.pem, domain.txt). The strings E:/CodeFiles/VulnScan/Vcenter/go_teamcheck/ and vcenter-db-teamcheck-go are also useful.

GTIG's BONESAW rule is the closest thing to a signature for this CVE. It targets exploit logic rather than campaign details: an ELF containing the vmdir database path, the SRP prime N, and both RPC interface UUIDs in ASCII or little-endian binary form.

rule G_Hunting_Utility_BONESAW_2 { meta: author = "Google Threat Intelligence Group" strings: $s_mdb_path = "/storage/db/vmware-vmdir/data.mdb" ascii $s_srp_prime_n = "AC6BDB41324A9A9BF166DE5E1389582FAF72B6651987EE07FC3192943DB56050" ascii $s_uuid_dbcp = "72c780fe-af6b-42ba-9b02-f61f4ccc98ed" ascii nocase $c_uuid_dbcp = { fe 80 c7 72 6b af ba 42 9b 02 f6 1f 4c cc 98 ed } $s_uuid_srp_verf = "c7e94609-6ab0-4767-8e92-b6485bff1649" ascii nocase $c_uuid_srp_verf = { 09 46 e9 c7 b0 6a 67 47 8e 92 b6 48 5b ff 16 49 } condition: uint32(0) == 0x464C457F and $s_mdb_path and $s_srp_prime_n and ($s_uuid_dbcp or $c_uuid_dbcp) and ($s_uuid_srp_verf or $c_uuid_srp_verf) }

The BEARBONES rule follows the same approach, keying on the banner and the /tmp host key paths:

rule G_Hunting_Backdoor_BEARBONES_1 { meta: author = "Google Threat Intelligence Group" strings: $c_sha256_len = { BA 0E 00 00 00 4C 89 E7 48 89 EE } $s_banner = "SSH-2.0-pulsesecure_" ascii $s_rsa_key = "/tmp/rsa_host_key" ascii $s_ecdsa_key = "/tmp/ecdsa_host_key" ascii $s_dss_key = "/tmp/dss_host_key" ascii condition: uint32(0) == 0x464C457F and ($c_sha256_len or $s_banner) and 2 of ($s_*) }

GTIG also published G_Hunting_Webshell_SLIMSAVAGE_1, G_Hunting_Webshell_CAGEMATCH_1, G_Hunting_Utility_SAWSNAKE_2 and G_Hunting_Backdoor_LAZYBONES_1. All six rules are in the Mandiant blog.

Behavioural signals on the staging host

Google SecOps customers get coverage through the Mandiant Frontline Threats, Emerging Threats and Hunt rule packs, under "Base64 Encoded Shell Command Execution", "Linux Web Server Process Echo Base64 Encoded String", "Suspicious Curl Network Connection to Staging Directory" and "Suspicious Curl Download to tmp". Other SIEMs can reproduce the logic: an unexpected SUID change on /bin/env, gdb -pid launched by a web server process, and curl writing executables into /tmp.

Caveat

Most of these IoCs (filenames, IPs, account names) are specific to this campaign and easy to change. The more durable detections are tied to how the exploit works: unexpected peers on TCP 2012, the RPC interface UUIDs, new SSO admins or machine accounts, and SSH being enabled through the VAMI API.

Affected systems and versions

Based on the Broadcom advisory response matrix and the NVD affected version data:

  • VMware vCenter 9.1.x: versions before 9.1.1.0200 (fixed in 9.1.1.0200)
  • VMware vCenter 9.0.x: versions before 9.0.2.0200 (fixed in 9.0.2.0200)
  • VMware vCenter 8.0: versions before 8.0 U3m (fixed in 8.0 U3m)
  • VMware vCenter 7.0: included in the advisory. Customers with an extended support contract should contact Broadcom Support. No generally available fixed build is listed in the response matrix excerpt.
  • VMware Cloud Foundation vCenter 5.x: versions before 5.2.5 (fixed in 5.2.5)
  • VMware Telco Cloud Platform: 3.0, 4.x, 5.0.x, 5.1.x and 5.2 (remediation per KB463116)
  • VMware Telco Cloud Infrastructure: 3.0 (remediation per KB463116)

The advisory also identifies VMware Cloud Foundation and VMware vSphere Foundation release branches, so check those product families as well as standalone vCenter. Broadcom states that no workaround is available. Confirm exact packaging and the supported update path against the Broadcom advisory and KB463116 before deployment.

References

Detect & fix
what others miss

Works with
  • GitHub
  • GitLab
  • Bitbucket
  • Azure DevOps Services
  • Gerrit
  • CVS
  • Jira
  • Linear
  • Slack
  • Security Compass
Explore integrations
Security magnifying glass visualization