A threat group suspected of ties to China has exploited a previously unknown vulnerability in VMware’s vCenter, the virtualization console, to gain unauthorized access to its targets. This vulnerability, which has been active since September, was patched on November 19, and Google has indicated that the console, which manages the entire virtualized data center, now requires additional hardening beyond the standard fix.
On November 19, Google’s Threat Intelligence Group (GTIG) and its subsidiary Mandiant detailed how this espionage group, dubbed ARMORY CASTLE, began exploiting an unpatched flaw in vCenter Server as early as the beginning of September.
This VMware console is responsible for configuring, managing, and monitoring ESXi hypervisors, effectively controlling all virtual machines within a data center. On the same day the vulnerability, identified as CVE-2026-123890, was disclosed, Broadcom, which owns VMware, released a critical patch that does not offer any workarounds.
Two Unknown Vulnerabilities and Unmonitored Appliances
The vulnerability allows attackers, without requiring a password, to retrieve a session key from vCenter’s directory service over port 2012, subsequently gaining access via the [email protected] account, the platform’s master account. Mandiant reports that the flaw arises from predictably generated credentials within the Secure Remote Password (SRP) authentication protocol, which exposes the shared key. According to Google, the exploit tool used by the group, named BONESAW by Mandiant, fully compromises vCenter appliances running versions 7 and 8. The metadata indicates its compilation date as July 13, just 16 days before Broadcom’s disclosure of a related flaw, CVE-2026-59309. This timing suggests that the group was aware of CVE-2026-123890 before the end of July, especially since BONESAW targeted a server already patched for the neighboring vulnerability in September.
ARMORY CASTLE additionally breached BeyondTrust’s Remote Support tool using another unknown vulnerability described by Google on October 15. From there, the group pivoted to internal Linux appliances lacking endpoint detection and response (EDR) tools. They deployed a remote access script (webshell) protected by the password “123,” elevated to superuser privileges, and scanned the network for VMware-specific ports. Their backdoor for the Secure Shell (SSH) administration protocol, derived from the open-source software Dropbear, was identified by a distinct banner: “SSH-2.0-pulsesecure_23345.” Researchers uncovered a network of compromised third-party firewalls and routers used to mask the origin of the connections by searching for this banner online.
From vCenter Directory to Virtual Machine Disks
Using access to these appliances, BONESAW extracted the vCenter directory database, retrieved the private key used to sign Security Assertion Markup Language (SAML) authentication tokens, and created an administrator account. A second tool, SAWSNAKE, forged tokens signed with this key, allowing the group to bypass multi-factor authentication and impersonate any user. They established a local administrator account on vCenter, enabled SSH access for the vCenter appliance through its management interface, stole the master encryption key of vCenter, and the credentials for its PostgreSQL database, subsequently closing off access immediately. Mandiant had previously observed the activation of SSH through the management interface during the BRICKSTORM intrusions, attributed to another group believed to be linked to China, UNC6201.
Finally, with superuser access on the ESXi hypervisors, the group directly read the virtual disks of a Windows server without ever booting the machine. They extracted the directory service account database from Active Directory and registry hives. Additionally, they employed a backdoor controlled via Slack messaging. “The use of three distinct sets of specialized malware and two unknown vulnerabilities during a single campaign indicates that ARMORY CASTLE is supported by significant resources,” note Google researchers.
System-Level Hardening for the Virtualization Console
“Given that vCenter serves as the centralized control plane for the virtualized data center, relying solely on post-fix patches is inadequate for its protection,” the researchers from Google assert. They pointed to the integrated firewall within the appliance, managed from the same interface that ARMORY CASTLE exploited. This firewall does not filter on a per-port basis, only allows incoming traffic, does not log rejected connection attempts, and an attacker can disable it upon accessing the interface.
Instead, Mandiant recommends a hardening tool that applies rules directly at the Linux kernel level of the appliance. This approach denies all by default, filters outgoing traffic, and sends events to the Security Information and Event Management (SIEM) tool. According to Google, these rules could have thwarted the attack at several stages, from port scanning to the use of forged tokens and data exfiltration, while also safeguarding against unknown vulnerabilities. Researchers recommend testing these rules and subjecting them to change control before applying them in production. Mandiant’s vSphere defense guide suggests dedicated administrative posts and configurations aligned with the Security Technical Implementation Guides (STIGs) from the U.S. Department of Defense.
An Emergency Patch and Exploitation Confirmed by Google
Broadcom has advised treating the patch as an emergency change per ITIL (Information Technology Infrastructure Library) guidelines, which are the benchmark for IT service management. The advisory also addresses four other vulnerabilities, three in SDDC Manager, the deployment tool for VMware Cloud Foundation, and one in VMware Tools for Windows. Notably, the FAQ section does not mention any exploitation details; information about active exploitation has only been corroborated by Google and Mandiant, who did not disclose the names of any victims.
For IT managers, this campaign shifts the focus to administrative consoles and internal appliances that may not always be covered by detection tools, as illustrated in this breach, yet can grant access to the entire virtualized environment. While Broadcom’s patch addresses the vulnerability, Mandiant’s proposed hardening aims to seal off potential pathways leading to the console.



