ARMORY CASTLE reached a VMware vCenter Server administrator session through a previously undisclosed directory-service flaw, then created an SSO administrator account and took a copy of the directory database. Google Threat Intelligence Group (GTIG) and Mandiant place the exploitation in early September, before the November 19 disclosure of CVE-2026-123890. The July fix for a different directory flaw left this route open; vCenter operators need the new patch and a compromise review.
The group had first compromised a BeyondTrust Remote Support appliance, then moved to an internal Linux appliance and toward the virtualization control system. vCenter manages ESXi hosts and their virtual machines, so a directory administrator session opens more than a single server. GTIG calls this a suspected PRC-nexus espionage operation; that attribution is GTIG’s assessment. The investigators observed administrator-account creation and database exfiltration in the examined intrusion.
A patched vCenter still fell
The attacker scanned internal systems for VMware-related services, including the vCenter SSO RPC interface on TCP 2012, then ran a Go utility GTIG calls BONESAW against vCenter. Mandiant’s logs place that execution in early September on a server already patched for CVE-2026-59309. The recovered utility has a July 13 build timestamp, though GTIG cautions that Go build metadata can be modified.
VMware’s record for the July flaw concerns an earlier authentication bypass in VMware Directory Service. GTIG describes that route as crafted LDAP authentication traffic on TCP 389 or 636. The newly disclosed flaw uses a different interface. Applying the July fix therefore left the observed RPC route available until its own patch.
RPC key grants SSO access
In VMware’s new CVE record, CVE-2026-123890 is an authentication bypass in VMware Directory Service, rated 9.8 under CVSS 3.1. GTIG describes an unauthenticated DCE/RPC Secure Remote Password session-key disclosure on TCP 2012: the rpc_srp_verifier service exposes a shared key, enabling an authenticated session as administrator@vsphere.local. That session reaches the vmdird RPC interface vmdirdbcp, which can be used to retrieve the directory’s data.mdb file. GTIG links the flaw to predictable UUIDv1 timestamps in the SRP interface.
BONESAW used that access to create an SSO administrator account and exfiltrate and parse the LMDB directory database, according to Mandiant. Its documented capabilities also include extracting the Security Token Service SAML signing key and forging administrative assertions. A companion utility, SAWSNAKE, can use those keys to obtain administrative session cookies and impersonate a user without that user’s MFA step. Treat a new SSO account or an unexpected database copy as a sign to investigate the identity system, even after the vCenter software is updated.
The foothold reached ESXi
On the intermediary Linux appliance, the group installed two JSP webshells: ActiveSer.jsp for command execution and u.jsp for staging files under /tmp. Mandiant observed privilege escalation there by changing SUID permissions on /bin/env, followed by staging of BONESAW at /tmp/123. The Linux appliances in this intrusion lacked endpoint detection and response tooling. On the first webshell, the parameter name was pwd and the hardcoded secret was 123; hunt for those tokens in web requests alongside the filename.
After the vCenter compromise, Mandiant observed the actor enable SSH via the VCSA management API path /rest/appliance/access/ssh, connect with a newly created administrator account, and take the vCenter master symmetric key and PostgreSQL credentials. The actor turned SSH off again. Later, the group authenticated to ESXi with vpxuser and root, put an SSH public key in root’s authorized keys, and staged material in a hidden .a directory on a VMFS volume. Mandiant observed the group extracting guest-VM files including Windows directory and registry material.
The group also ran a modified SSH server, BEARBONES, on the Linux appliance; its distinctive banner was SSH-2.0-pulsesecure_23345. Separately, Mandiant found a Slack-based LAZYBONES backdoor at /usr/bin/vmtoolscs on another Linux appliance, persisted through a systemd service that restarted every 50 minutes. The backdoor abused the legitimate shared service slack.com; prioritize host and channel context over blocking the entire platform.
Patch the directory-service route
VMware’s CVE record lists the first fixed vCenter builds as 9.1.1.0200 for 9.1.x, 9.0.2.0200 for 9.0.x, 8.0 U3m for 8.0, and 5.2.5 for the listed 5.x line. It also lists Telco Cloud Platform 3.0, 4.x, 5.0.x, 5.1.x and 5.2, plus Telco Cloud Infrastructure 3.0, as affected; operators of those products should use Broadcom’s product-specific update guidance. GTIG describes BONESAW as targeting vCenter 7.x and 8.x in its research; ask Broadcom for a supported migration or remedy for a 7.x estate rather than treating an 8.x build as its patch. GTIG recommends that affected organizations “immediately apply the security patches provided by Broadcom.”
Restrict TCP 2012 to approved administration sources at a network control outside the appliance, and restrict the vCenter UI and API on TCP 443 and VAMI on 5480 to the same managed access path where operations allow it. An external filter stops a host on an unapproved network from reaching those vCenter management ports while the update is installed. GTIG warns that the native VAMI GUI firewall lacks per-port precision and remote denied-connection logging, and an actor with VAMI access can alter it. Test the allowlist and remote logging under change control before enforcing default deny; blocking a legitimate vCenter management path can interrupt operations.
Verify the vCenter identity state
Inventory the installed vCenter build, apply the branch-specific fix, then verify the running build against the fixed version. From an ordinary workload network, test whether TCP 2012 remains reachable; expected result is a block unless that source is explicitly approved. From a managed administration source, confirm the needed access still works and the deny logs reach the remote collector.
Review SSO directory account creation and privilege assignments for unfamiliar administrators, including accounts with the sps- prefix in the vsphere.local SSO domain; inspect VAMI SSH-enable events at /rest/appliance/access/ssh, SSH access to vCenter and ESXi, access to /storage/db/vmware-vmdir/data.mdb, and activity around the vCenter STS signing key. On the intermediary appliances, hunt for ActiveSer.jsp, u.jsp, /tmp/123, /tmp/sss, /tmp/dbr, and /usr/bin/vmtoolscs. Examine ESXi root authorized keys and unexpected VMFS .a staging directories. GTIG also publishes a YARA rule set for these tool families in its investigation. Its suspected obfuscation-network IP indicators are 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, and 81[.]45[.]89[.]43. Search historical connections before imposing a block on third-party infrastructure. An isolated match needs context: correlate accounts, host access and file activity across the BeyondTrust appliance, Linux systems, vCenter and ESXi.
If the SSO directory or signing key was taken, treat identity recovery as part of incident response alongside the version upgrade: preserve evidence, remove unauthorized accounts and access paths, and follow Broadcom’s supported procedure to rotate affected credentials and signing material.