A suspected China-linked espionage group broke into a victim's VMware vCenter server with a flaw nobody else knew about, created itself an administrator account, and used that foothold to copy Active Directory secrets straight off the disks of virtual machines. Mandiant and Google Threat Intelligence Group (GTIG) published the details on November 19, the same day VMware disclosed the bug, tracked as CVE-2026-123890.
The actor is ARMORY CASTLE, the group Google formerly tracked as UNC5221. In October, Google reported the same crew exploiting a zero-day in BeyondTrust Remote Support (our earlier coverage). The new report picks up where that one stopped: what the attackers did after they were inside.
A login that needs no password
vCenter is the console that manages fleets of ESXi hypervisors, which makes it one of the most valuable boxes in a data centre. According to GTIG, CVE-2026-123890 lets an unauthenticated attacker recover a session key from vCenter's directory service over its RPC interface on TCP port 2012, which produces an authenticated session as administrator@vsphere[.]local without knowing any password. Google traces the root cause to predictable UUIDv1 timestamp generation in the Secure Remote Password (SRP) authentication interface.
ARMORY CASTLE weaponised it with a Go tool Mandiant calls BONESAW. GTIG says BONESAW streams out vCenter's internal directory database, extracts the Security Token Service SAML signing key, creates a Single Sign-On administrator, and grants Global Administrator rights on the vCenter root folder. Google says it works against vCenter 7.x and 8.x.
The timeline is the striking part. Logs show BONESAW ran in early September 2026, against a server that was already patched for a related vCenter flaw, CVE-2026-59309, disclosed on July 29. The recovered binary carries a build timestamp of July 13, which GTIG notes can be tampered with, but Mandiant says it has evidence the group knew about CVE-2026-123890 before CVE-2026-59309 was made public.
From one appliance to the domain controllers
The intrusion started on the compromised BeyondTrust appliance. ARMORY CASTLE scanned for VMware ports, then moved to an internal Linux appliance with no endpoint detection, where it dropped two webshells (SLIMSAVAGE for commands, CAGEMATCH for file uploads) and gained root by setting the SUID bit on /bin/env with the gdb debugger.
After BONESAW, a second tool, SAWSNAKE, forged SAML tokens for vCenter, a so-called Golden SAML attack that lets the holder impersonate any user and sidestep MFA. A backdoored Dropbear SSH server, BEARBONES, accepted a hard-coded password and logged real ones. Its banner, SSH-2.0-pulsesecure_23345, led Google to what it suspects is a relay network of hacked third-party firewalls and routers.
Inside vCenter, the group briefly enabled SSH, took the master encryption key and PostgreSQL credentials, then switched SSH off again. On the ESXi hosts it used a Python toolkit to read guest virtual disks directly and pull out the NTDS database and the SYSTEM and SECURITY registry hives from a Windows machine, which together hold a domain's password hashes. Google says the campaign used two zero-days and three sets of purpose-built malware, a sign of deep platform expertise and substantial resources.
Apply Broadcom's vCenter patch now
GTIG strongly recommends applying Broadcom's security update for CVE-2026-123890 immediately. Defenders should also restrict network access to vCenter's port 2012, look for unexpected SSO administrator accounts and VAMI API calls that toggle SSH, and hunt for the SSH banner SSH-2.0-pulsesecure_23345 on non-standard high ports. Network indicators GTIG published include:
- 103[.]147[.]251[.]5
- 118[.]174[.]128[.]34
- 176[.]9[.]89[.]66
- 185[.]71[.]15[.]109
- 81[.]45[.]89[.]43
The full list, detections and hardening guidance are in the original report by Jayson Frost, Bryce Greathouse, Dhanesh Kizhakkinan, John Palmisano, Steve Rasch, Daniel Sislo and John Wolfram. The lesson for anyone running vSphere is uncomfortable: the attackers moved through appliances that carry no EDR agent, where nobody was watching.
This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.