Broadcom has issued an emergency security advisory patching a maximum-severity vulnerability in VMware vCenter Server that allows remote, unauthenticated attackers to seize complete administrative control over virtualized enterprise environments. Tracked as CVE-2026-123890, the flaw carries a Common Vulnerability Scoring System (CVSS v3.1) base score of 9.8 out of 10. It stems from a critical authentication bypass in the VMware Directory Service (vmdir), the central identity and access management component underpinning VMware’s vSphere virtualization platform.
The security update, cataloged under advisory VMSA-2026-0009 on November 19, 2026, took on heightened urgency when threat intelligence researchers at Google Cloud’s Mandiant disclosed that sophisticated state-sponsored actors had been actively weaponizing the flaw as a covert VMware vCenter zero-day since at least early September 2026.
Forensic analysts attributed the espionage campaign to ARMORY CASTLE—a prolific advanced persistent threat (APT) group with suspected ties to the People’s Republic of China, previously tracked as UNC5221 and Warp Panda. The threat actor leveraged the flaw to breach high-value corporate networks, execute custom malware, forge Single Sign-On (SSO) authentication tokens, and siphon sensitive virtual machine disks directly from physical host servers without triggering traditional software alerts.
Broadcom confirmed that there are no viable software workarounds or configuration mitigations. In an accompanying technical notice, the vendor classified the patch deployment as an emergency change under Information Technology Infrastructure Library (ITIL) operational standards, advising IT and security leaders worldwide to accelerate maintenance windows and patch immediately.
"We investigated multiple intrusions involving ARMORY CASTLE's exploitation of this vulnerability prior to vendor disclosure," stated Charles Carmakal, Chief Technology Officer at Mandiant, confirming that the attackers breached internal systems weeks before public awareness. "Organizations cannot simply patch and assume safety. A comprehensive compromise assessment is required."
+------------------------------------------------------------------------------------+
| CVE-2026-123890 AT A GLANCE |
+----------------------+-------------------------------------------------------------+
| Severity Rating | Critical (CVSS v3.1 Base Score: 9.8) |
+----------------------+-------------------------------------------------------------+
| Flaw Type | Authentication Bypass Using Alternate Path (CWE-288) |
+----------------------+-------------------------------------------------------------+
| Vulnerable Subsystem | VMware Directory Service (vmdir) / DCE/RPC Interface |
+----------------------+-------------------------------------------------------------+
| Network Vectors | TCP Port 2012 (Direct vmdir RPC listener) |
+----------------------+-------------------------------------------------------------+
| Threat Actor | ARMORY CASTLE (UNC5221 / Warp Panda / China-nexus) |
+----------------------+-------------------------------------------------------------+
| Attack Timeline | Zero-day exploitation: Sept 2026; Public Patch: Nov 19, 2026|
+----------------------+-------------------------------------------------------------+
| Workarounds | None available; Requires binary upgrade and compromise audit|
+----------------------+-------------------------------------------------------------+
The Nervous System of Enterprise Computing: Why vCenter Matters
To grasp why a vulnerability in VMware vCenter Server triggers high-priority operational alarms across global security operations centers, one must understand how modern data centers are structured.
For over two decades, enterprise computing has run on virtualization. Rather than purchasing, cabling, powering, and cooling a separate physical server for every individual software application—such as an enterprise resource planning database, a customer portal, or an internal payroll service—organizations use virtualization to divide one physical computer into dozens or hundreds of independent, isolated digital computers known as virtual machines (VMs).
This architectural trick is performed by a hypervisor—a thin, highly optimized software layer installed directly onto bare physical computer hardware. In VMware’s software ecosystem, that bare-metal hypervisor is called VMware ESXi. ESXi commands the physical processors (CPUs), system memory (RAM), storage drives, and network interfaces, carving them up into slices and presenting them to guest operating systems like Microsoft Windows Server or Red Hat Enterprise Linux.
Managing one ESXi host is simple through a local web console. However, a Fortune 500 bank, a healthcare network, a defense contractor, or a major telecommunications operator does not run a single physical host; they operate hundreds or thousands of them distributed across multiple continents, supporting hundreds of thousands of active virtual machines.
That is where VMware vCenter Server enters the architecture.
+---------------------------------------+
| VMware vCenter Server |
| (Centralized Management Plane) |
| |
| +---------------------------------+ |
| | VMware Directory Service | |
| | (vmdir - Identity & SSO Engine)| |
| +---------------------------------+ |
+-------------------+-------------------+
|
+-------------------------------+-------------------------------+
| (Encrypted Management Network)| | (Encrypted Management Network)
v v v
+------------------+ +------------------+ +------------------+
| ESXi Host A | | ESXi Host B | | ESXi Host C |
| (Bare-Metal OS) | | (Bare-Metal OS) | | (Bare-Metal OS) |
| | | | | |
| [VM1] [VM2] | | [VM3] [VM4] | | [VM5] [VM6] |
| (Payroll) (Active| | (Customer (Prod | | (Core IP (R&D |
| Directory) | Portal) SQL) | | Router) Database|
+------------------+ +------------------+ +------------------+
vCenter Server acts as the centralized brain, control plane, and orchestrator of this entire virtualization empire. Through vCenter, a systems administrator can:
- Spin up, configure, modify, back up, or terminate any virtual machine across any data center worldwide.
- Live-migrate running virtual machines between physical servers across cities with zero application downtime using VMware vMotion.
- Control virtual firewalls, software-defined networks (VMware NSX), and virtual storage area networks (vSAN).
- Take complete snapshots of running servers, creating an exact freeze-frame copy of a computer’s disk, operating system state, and active memory contents.
Whoever controls vCenter Server effectively commands the entire virtual computing infrastructure of the organization. If an intruder compromises a single virtual machine running a web application, that intruder is trapped inside that individual workload’s sandbox. But if an intruder gains administrative access to vCenter Server, the sandbox disappears. The attacker sits above the operating systems, possessing master authority over all hosts, all networks, and all stored data.
Deconstructing CVE-2026-123890: The Flaw Inside the Directory
The 9.8 CVSS score assigned to CVE-2026-123890 reflects the worst possible profile in vulnerability metrics: it is remotely exploitable over a network, requires zero initial privileges, requires zero human interaction, and results in total loss of confidentiality, integrity, and operational availability.
Understanding how an unauthenticated network packet transforms into full administrative control requires peeling back the internal plumbing of vCenter’s authentication engine.
The Role of vmdir and Single Sign-On
Inside the vCenter Server virtual appliance (VCSA)—a hardened Linux-based software appliance running VMware’s Photon OS—runs a core daemon called vmdird, short for the VMware Directory Service.
The Directory Service is a specialized lightweight directory access protocol (LDAP) and authentication service. It acts as the authoritative store for all internal identities, permissions, groups, and Single Sign-On (SSO) configurations across the vSphere management domain. When an administrator attempts to log into the vCenter web interface at https://
vmdir holds the master cryptographic keys, the domain user database, and the root administrative account: [email protected].
The DCE/RPC Communication Vector
To synchronize data between multiple vCenter instances (such as environments utilizing Enhanced Linked Mode) and allow internal subcomponents to coordinate, vmdir exposes a legacy communication protocol known as Distributed Computing Environment / Remote Procedure Calls, or DCE/RPC.
DCE/RPC is a foundational network architecture developed in the early 1990s that enables a computer program on one machine to execute procedures or subroutines on another machine without the programmer explicitly writing the networking code. In vCenter, vmdir listens for remote procedure calls directly over TCP port 2012.
Historically, network services that communicate over DCE/RPC have suffered from memory safety issues, buffer overruns, and input sanitization failures. But CVE-2026-123890 is not a standard memory corruption bug; it is an authentication bypass involving alternate communication channels, cataloged under Common Weakness Enumeration CWE-288.
VULNERABILITY EXECUTION FLOW (CVE-2026-123890)
[ Remote Adversary ]
|
| 1. Sends malformed RPC request over TCP Port 2012
v
+------------------------------------------------------------------------+
| vCenter Server Appliance (VCSA) |
| |
| +------------------------------------------------------------------+ |
| | VMware Directory Service Daemon (vmdird) | |
| | | |
| | * Intercepts incoming DCE/RPC call | |
| | * Evaluates SRP-6a handshaking logic | |
| | * CWE-288 Logic Flaw: Bypasses authentication challenge checks | |
| | * Leaks session validation material / grants master admin token | |
| +---------------------------------+--------------------------------+ |
| | |
| v |
| +------------------------------------------------------------------+ |
| | Identity Established: [ [email protected] ] | |
| | Access: Full root-level management plane control | |
| +------------------------------------------------------------------+ |
+------------------------------------------------------------------------+
The Cryptographic Failure: SRP-6a and Session Key Disclosures
Security analysts tracing the vulnerability discovered that the weakness lies in how vmdir implements the Secure Remote Password (SRP-6a) protocol over its RPC interface.
SRP-6a is an augmented password-authenticated key exchange (PAKE) protocol. In computer science, SRP is designed to solve a fundamental security paradox: how can a client prove to a server that it knows a secret password without ever sending the password—or even a hash of the password—across the network?
Under a standard SRP protocol exchange:
- The user provides an identity (username).
- The server responds with a random number called a cryptographic salt and public parameters.
- Both client and server perform modular arithmetic calculations using their respective public and private ephemeral keys.
- If both calculations match, a strong, shared session key is derived, mathematically confirming that the client knows the password without exposing the credential to network eavesdroppers.
In the case of CVE-2026-123890, the implementation of this cryptographic negotiation inside vmdird contained a catastrophic logic omission. When receiving an intentionally crafted sequence of remote procedure call requests over TCP port 2012, the service improperly handled internal state transitions.
Instead of terminating the connection when an unauthenticated caller failed to provide valid client-proof parameters, the service either validated the session prematurely or disclosed the derived SRP session key back across the network channel.
Armed with this key, the calling program is immediately acknowledged by vmdir as an authenticated administrative peer. The attacker does not need to guess credentials, run brute-force wordlists, or crack password hashes. By sending a specific sequence of network frames to TCP port 2012, the external attacker is instantly handed an authenticated session operating under the authority of [email protected].
Anatomy of an Intrusion: The ARMORY CASTLE Weaponization Chain
Vulnerabilities discovered in academic research labs are theoretical threats until threat actors operationalize them. What elevates CVE-2026-123890 into a national-security-grade crisis is the forensic trail left by the threat group that weaponized it in the wild.
The threat group responsible is ARMORY CASTLE. Tracked across the cybersecurity industry under various designations—including UNC5221 by Google Cloud's Mandiant and Warp Panda by CrowdStrike—this cyberespionage cluster is notorious for executing targeted campaigns against critical infrastructure, defense organizations, telecommunications providers, and technology vendors.
Unlike opportunistic cybercriminals who blast automated ransomware across the open internet, ARMORY CASTLE specializes in living off the edge: finding zero-day flaws in enterprise perimeter gateways, VPN concentrators, and internal infrastructure appliances, operating in the shadows for months without generating standard antivirus alarms.
ARMORY CASTLE ATTACK KILL CHAIN
+-------------------------------------------------------------------------+
| STEP 1: EDGE ACCESS |
| Exploit BeyondTrust Remote Support Appliance (CVE-2026-102764) |
| Attacker breaches external perimeter and establishes internal foothold. |
+------------------------------------+------------------------------------+
|
v
+-------------------------------------------------------------------------+
| STEP 2: INTERNAL RECONNAISSANCE & LATERAL MOVEMENT |
| Pivot through internal subnets; scan for VMware management endpoints. |
| Locate vCenter Server listening on TCP Ports 2012, 443, and 5480. |
+------------------------------------+------------------------------------+
|
v
+-------------------------------------------------------------------------+
| STEP 3: ZERO-DAY EXPLOITATION |
| Deploy BONESAW utility over TCP Port 2012 (CVE-2026-123890). |
| Bypass authentication; gain [email protected] session. |
+------------------------------------+------------------------------------+
|
v
+-------------------------------------------------------------------------+
| STEP 4: DATABASE & KEY EXTRACATION |
| Exfiltrate Lightning Memory-Mapped Database (LMDB). |
| Steal master encryption keys and SAML token-signing x509 certificates. |
+------------------------------------+------------------------------------+
|
v
+-------------------------------------------------------------------------+
| STEP 5: PERSISTENCE & FORGERY |
| Create shadow SSO administrator account (e.g., "sps-admin"). |
| Deploy SAWSNAKE to execute Golden SAML attack, minting valid tokens |
| that completely bypass Multi-Factor Authentication (MFA). |
+------------------------------------+------------------------------------+
|
v
+-------------------------------------------------------------------------+
| STEP 6: HYPERVISOR ESCALATION & DATA THEFT |
| Pivot from vCenter down into bare-metal ESXi hypervisors. |
| Mount Windows guest virtual machine disks (VMDKs) directly from ESXi. |
| Extract confidential enterprise databases without guest OS visibility. |
+-------------------------------------------------------------------------+
Forensic timelines assembled by incident responders trace the complete life cycle of an ARMORY CASTLE intrusion.
Phase 1: Ingress via Perimeter Gateways
ARMORY CASTLE did not begin their assault by knocking directly on vCenter’s front door. In enterprise networks built according to basic security standards, vCenter is tucked deep inside an internal management network and should never be exposed directly to the public internet.
To reach it, the attackers first breached an external network edge device. In early September 2026, ARMORY CASTLE leveraged an initial zero-day vulnerability in an external remote support system—specifically a flaw tracked as CVE-2026-102764 affecting BeyondTrust Remote Support appliances. Once inside the compromised appliance, the threat actors used its internal network connections to pivot laterally into the victim’s core corporate network.
Phase 2: Unleashing BONESAW Against vCenter
Moving silently through the internal subnet, the actors mapped out management appliances and targeted the vCenter Server deployment. To execute the breach, ARMORY CASTLE utilized a custom, compiled Golang utility named BONESAW.
BONESAW was custom-built to interact with vCenter’s proprietary Directory Service interfaces. The tool initiated a connection over TCP port 2012, triggered CVE-2026-123890, exploited the SRP key disclosure flaw, and instantly established an unauthenticated session with full administrative privileges.
The weaponization of this VMware vCenter zero-day through BONESAW gave ARMORY CASTLE instantaneous root-level dominance over the central virtualization server.
Phase 3: The LMDB Heist and Cryptographic Plunder
Once authenticated, BONESAW did not simply look around; it executed a targeted cryptographic theft.
The utility reached into the vCenter appliance’s local storage and extracted the Lightning Memory-Mapped Database (LMDB). LMDB is an ultra-fast, zero-copy, key-value embedded database used by vmdird to store directory records. Inside this database resides the cryptographic foundation of the enterprise:
- The Master Token-Signing Key: The private x509 cryptographic certificate that vCenter’s Identity Provider uses to digitally sign Security Assertion Markup Language (SAML) tokens.
- vCenter Master Encryption Keys: The cryptographic keys used to unlock encrypted virtual machines and protect stored virtual hardware secrets.
- Internal Appliance Credentials: Cleartext or reversibly encrypted passwords used by internal services to communicate with underlying storage pools and ESXi hosts.
Phase 4: Golden SAML and SAWSNAKE
Having stripped vCenter of its signing certificates, ARMORY CASTLE deployed a secondary post-exploitation module dubbed SAWSNAKE.
SAWSNAKE weaponized the stolen x509 certificate to execute a technique known as a Golden SAML attack. In modern identity architecture, when an authorized user successfully passes a login check (including entering passwords and clearing smartphone-based multi-factor authentication prompts), the identity system generates a cryptographically signed SAML assertion ticket. Every other system in the network trusts this ticket because it bears the mathematical signature of the identity provider’s private key.
By stealing that private key, SAWSNAKE could mint authentic-looking SAML tickets for any identity it chose. The threat actors forged credentials on demand, granting themselves administrative sessions that bypassed Multi-Factor Authentication (MFA) entirely. To ensure long-term persistence, they also leveraged the API to create rogue local administrator accounts, often disguised under legitimate-looking service account names such as sps-admin.
Phase 5: Hypervisor Escalation and Silent Data Exfiltration
With total control of vCenter, ARMORY CASTLE shifted downward toward the physical hypervisors.
Using vCenter's administrative APIs, the threat actors pushed their own public SSH keys onto the root accounts of physical ESXi servers. They connected to the bare-metal ESXi consoles, completely bypassing the guest operating systems running on top of them.
From the bare-metal ESXi hypervisor, the attackers directly mounted the virtual disk files (.vmdk) belonging to critical Windows virtual machines—such as file servers and corporate databases. They cloned and extracted these files without ever logging into Windows, without triggering a single Windows Event Log, and without alerting the modern Endpoint Detection and Response (EDR) software installed inside those virtual machines.
Why vCenter Exploits Keep Recurring: The Legacy Code Burden
CVE-2026-123890 is not an isolated lightning strike. It is the latest chapter in an ongoing security challenge surrounding VMware’s management plane.
A historical review of high-profile vulnerabilities reveals that vCenter Server's core identity and remote procedure call components have repeatedly been targeted by advanced threat actors:
+----------------------------------------------------------------------------------------------------+
| CHRONOLOGY OF CRITICAL VCENTER SERVER VULNERABILITIES |
+-----------------+---------------+-----------+------------------------------------------------------+
| CVE Identifier | Patch Date | CVSS Base | Technical Vulnerability Nature |
+-----------------+---------------+-----------+------------------------------------------------------+
| CVE-2023-34048 | October 2023 | 9.8 | Out-of-bounds write in DCERPC implementation; |
| | | | exploited by UNC3886 as a zero-day since 2021. |
+-----------------+---------------+-----------+------------------------------------------------------+
| CVE-2024-38812 | Sept/Nov 2024 | 9.8 | Heap-based buffer overflow in DCE/RPC handling; |
| | | | led to unauthenticated remote code execution. |
+-----------------+---------------+-----------+------------------------------------------------------+
| CVE-2026-59309 | July 2026 | 9.8 | vmdir authentication bypass and directory traversal; |
| | | | exploited in the wild within 5 days of release. |
+-----------------+---------------+-----------+------------------------------------------------------+
| CVE-2026-123890 | November 2026 | 9.8 | Authentication bypass via SRP-6a session leak in |
| | | | vmdir; exploited by ARMORY CASTLE as zero-day. |
+-----------------+---------------+-----------+------------------------------------------------------+
Why do these severe flaws cluster around the exact same components? The answer lies in the engineering history of enterprise software.
vCenter Server is not a lightweight, modern microservice built from scratch in recent years. It is an enterprise platform developed over more than two decades. Beneath modern HTML5 user interfaces and RESTful APIs lies millions of lines of foundational C and C++ code, built upon legacy distributed networking frameworks such as DCE/RPC and customized LDAP directory structures.
VCENTER SERVER ARCHITECTURAL LAYERS
[ Modern Web UI (HTML5) ] [ REST APIs (VAMI/vSphere) ]
| |
+--------------+--------------+
|
v
+-----------------------------------------+
| vSphere Management Service |
+--------------------+--------------------+
|
v
+-----------------------------------------+
| VMware Directory Service (vmdir) |
| |
| [ Legacy DCE/RPC Engine (TCP 2012) ] | <-- ATTACK SURFACE
| [ SRP-6a Key Negotiation Protocols ] | (CVE-2023-34048,
| [ Embedded LMDB Database Storage ] | CVE-2026-59309,
+-----------------------------------------+ CVE-2026-123890)
DCE/RPC protocols were designed in an era when enterprise networks were assumed to be fundamentally trustworthy. They incorporate complex data serialization models, internal pointer-swapping, and multi-step state machine negotiations.
When modern security engineers audit software, they frequently discover that legacy network listeners:
- Fail to enforce strict bounds checking on dynamic memory buffers, leading to heap overflows.
- Contain logic flaws in state transitions where unauthenticated packets are evaluated before authorization gates close.
- Expose management endpoints to network segments that should never have direct access to internal directory structures.
Compounding this engineering challenge, Mandiant's incident response team noted that ARMORY CASTLE possessed technical working knowledge of CVE-2026-123890 before Broadcom patched the earlier July flaw, CVE-2026-59309. Threat groups conduct their own reverse-engineering of vendor updates. When a vendor patches one flaw in vmdir, foreign intelligence researchers study the patch diffs, dissect surrounding functions in the source code, and identify adjacent edge cases that the vendor's engineers overlooked.
The Economics of Hypervisor Targeting: The EDR Blind Spot
From an intelligence agency's perspective, spending top-tier zero-day exploits on hypervisor infrastructure offers unmatched return on investment.
In corporate security, organizations have spent billions of dollars deploying Endpoint Detection and Response (EDR) software. Agents from vendors such as CrowdStrike, Microsoft Defender, and SentinelOne run on every employee laptop and virtual server. These agents inspect running memory, record process creation, analyze network sockets, and use machine learning to block unauthorized activity.
However, the hypervisor management plane represents an architectural blind spot.
THE HYPERVISOR VISIBILITY DIVIDE
+------------------------------------------------------------------------+
| GUEST VIRTUAL MACHINE |
| |
| [ Corporate Database / ERP Software ] |
| [ Endpoint Detection & Response (EDR) Agent ] |
| ^ |
| |-- EDR sees everything INSIDE this operating system. |
+---------+--------------------------------------------------------------+
|
==========|=============================================================== VISIBILITY DIVIDE
| (Virtual Hardware Boundary)
+---------+--------------------------------------------------------------+
| BARE-METAL HYPERVISOR (ESXi) & MANAGEMENT PLANE (vCenter) |
| |
| * vCenter Appliance (Photon OS - Black-box Linux environment) |
| * Physical ESXi Hypervisor (VMkernel bare-metal architecture) |
| |
| [ ZERO TRADITIONAL EDR AGENTS INSTALLED ] |
| |
| Attacker sits here: Extracts raw disk blocks, inspects memory states, |
| and clones workloads without the Guest OS or its EDR ever knowing. |
+------------------------------------------------------------------------+
Because vCenter Server is distributed as an integrated, hardened virtual appliance, administrators cannot easily install third-party host-based security agents inside it. Physical ESXi servers run on a proprietary microkernel operating system called the VMkernel, which intentionally lacks support for standard third-party background software agents.
Consequently, if an adversary attacks an operating system from the inside, alarms sound instantly. But if an adversary attacks from the hypervisor plane:
- No Process Creation Alerts: The virtual machine’s operating system never registers that a rogue process ran, because the commands executed in the hypervisor outside its virtual boundary.
- No File Access Violations: When the attacker reads the virtual disk directly from ESXi storage, the guest file system's access-control lists (ACLs) are bypassed.
- Long Operational Dwell Times: Attackers maintain access inside hypervisor networks for months or years without discovery.
This dynamic explains why state-sponsored actors dedicate deep resources to discovering zero-days in virtualization infrastructure. The discovery of another critical VMware vCenter zero-day underscores how the underlying management layer has become a primary target for sophisticated cyber warfare campaigns.
Upgrade Complexities and Enterprise Friction: The Patching Dilemma
When Broadcom published VMSA-2026-0009 on November 19, systems administrators encountered an immediate dilemma: deploying the security update was operationally complex.
Patching enterprise virtualization software is not as simple as clicking "Update" on a personal phone. vCenter Server is deeply integrated into operational workflows. Applying updates can disrupt automated backups, pause disaster-recovery replication, and temporarily take down management access across entire server farms.
Broadcom's advisory introduced several specific technical and operational challenges that complicate remediation.
+---------------------------------------------------------------------------------------------+
| AFFECTED PLATFORMS AND REMEDIATION |
+----------------------------------+-----------------------+----------------------------------+
| Product Family | Affected Baseline | Minimum Fixed Release / Build |
+----------------------------------+-----------------------+----------------------------------+
| VMware vCenter Server 8.0 | 8.0 prior to U3m | 8.0 Update 3m |
+----------------------------------+-----------------------+----------------------------------+
| VMware vCenter Server 9.0 | 9.0 prior to 9.0.2 | 9.0.2.0200 |
+----------------------------------+-----------------------+----------------------------------+
| VMware vCenter Server 9.1 | 9.1 prior to 9.1.1 | 9.1.1.0200 |
+----------------------------------+-----------------------+----------------------------------+
| VMware Cloud Foundation (VCF) | 5.x Deployments | 5.2.5 Async Patch |
+----------------------------------+-----------------------+----------------------------------+
| VMware Telco Cloud Platform | 3.x, 4.x, 5.x | Contact Support / Apply KB Matrix|
+----------------------------------+-----------------------+----------------------------------+
| VMware vCenter Server 7.0 | All 7.0 Releases | Restricted to Extended Contracts |
+----------------------------------+-----------------------+----------------------------------+
The "Quick Patch" Incompatibility
vCenter includes a feature called Quick Patch, designed to let systems administrators pull down targeted, lightweight micro-updates to patch vulnerabilities rapidly without performing a heavy, full-system upgrade.
In advisory VMSA-2026-0009, Broadcom warned that CVE-2026-123890 cannot be resolved using Quick Patch. The updates are cumulative binary replacements of fundamental infrastructure services. Organizations must stage full appliance updates, requiring complete service restarts, staging reboots, and formal IT change windows.
The Cloud Foundation "Back-in-Time" Upgrade Block
Organizations running VMware Cloud Foundation (VCF)—VMware’s flagship unified software suite bundling compute, storage, and networking into a single deployment—ran straight into an architectural upgrade conflict.
Broadcom’s FAQ confirmed that if an engineering team running VCF applied the standalone vCenter 8.0 Update 3m or vCenter 9.0.2.0200 patch to secure their systems immediately, they would trigger a software version conflict known as a "Back-in-Time" restriction. Once applied, the underlying vCenter build version becomes higher than the vCenter build packaged inside the broader VMware Cloud Foundation 9.1.1 release bundle, preventing future platform upgrades until Broadcom issues VCF 9.1.2.
Infrastructure architects were left choosing between leaving their core servers exposed to active zero-day exploitation or applying an emergency patch that disrupts their long-term enterprise upgrade roadmap.
The Legacy vCenter 7.0 Crisis
The sharpest point of contention centers on legacy platforms. Millions of virtual machines worldwide still run on clusters managed by VMware vCenter Server 7.0. vCenter 7.0 reached its official End of General Support (EOGS) in October 2025.
Under Broadcom’s business model and support guidelines, public security patches are no longer posted for unsupported branches. Organizations running vCenter 7.0 cannot download a public patch from Broadcom's support portal. Instead, Broadcom's advisory directs them to contact customer support—and access to backported fixes is restricted to enterprises that have purchased expensive extended support contracts.
This leaves underfunded public sector entities, municipal utilities, hospital networks, and smaller businesses that have not migrated to vSphere 8.0 or 9.0 exposed to an actively exploited threat without an immediate, accessible patch.
Technical Incident Response Playbook: Hunting for Compromise
Because the VMware vCenter zero-day was weaponized prior to public patch availability, organizations cannot assume their systems are secure simply because they installed the patch. If an intrusion occurred in September or October 2026, applying the patch today merely locks the front door while the intruder sits comfortably inside with stolen keys.
Every organization running affected vCenter systems must execute a dual-track response: apply the patch to prevent future exploitation, and launch an immediate forensic investigation to detect past compromise.
ENTERPRISE RESPONSE DUAL-TRACK PLAYBOOK
+-----------------------------------+
| CVE-2026-123890 Incident Action |
+-----------------+-----------------+
|
+-------------------------+-------------------------+
| |
v v
+-------------------------------+ +------------------------------------+
| TRACK 1: HARDENING & PATCHING | | TRACK 2: COMPROMISE INVESTIGATION |
| | | |
| 1. Apply 8.0 U3m / 9.0.2 / | | 1. Audit /var/log/vmware/vmdir/ |
| 9.1.1 cumulative updates. | | for anomalous SRP sessions. |
| 2. Restrict TCP Port 2012 | | 2. Query SSO database for rogue |
| via perimeter firewalls. | | admin accounts (sps-* patterns).|
| 3. Isolate management plane | | 3. Inspect VAMI API access logs |
| to out-of-band VLANs. | | for transient SSH activation. |
| 4. Terminate direct Internet | | 4. Sweep ESXi authorized_keys for |
| ingress paths. | | unauthorized root SSH certs. |
+-------------------------------+ +-----------------+------------------+
|
v
+------------------------------------+
| RECOVERY IF COMPROMISE DETECTED |
| |
| * Revoke & re-mint SAML certs. |
| * Rotate master encryption keys. |
| * Rebuild vCenter from golden ISO. |
| * Force global credential resets. |
+------------------------------------+
Forensic Check 1: Auditing the Directory Service for Rogue Accounts
Security teams should inspect the vCenter Single Sign-On user database for accounts created outside of authorized IT change requests.
Administrators can connect to the vCenter appliance shell over an out-of-band administrative console and list all identity accounts within the vsphere.local domain:
# Query all accounts in the internal directory service
/usr/lib/vmware-vmafd/bin/dir-cli user find-by-name --level 2
# Inspect members of the top-level Administrator group
/usr/lib/vmware-vmafd/bin/dir-cli group listusers --name Administrators
Forensic Marker: Threat analysts specifically observed ARMORY CASTLE provisioning rogue administrative accounts designed to mimic legitimate vCenter background services. Look for newly created accounts starting with sps-, such as sps-admin, sps-service, or sps-maintenance. In a clean deployment, the sps (Storage Policy Server) uses internal token exchanges and does not possess a standalone administrative account inside the directory.
Forensic Check 2: Inspecting VAMI API Logs for SSH Toggling
Because permanent external SSH access to vCenter is easily noticed by automated security audits, ARMORY CASTLE used vCenter’s REST API to turn SSH on, execute commands, and shut it down minutes later.
Examine the vCenter Server Appliance Management Interface (VAMI) web access logs:
# Search for access to the appliance SSH configuration endpoint
grep -E "POST /rest/appliance/access/ssh" /var/log/vmware/applmgmt/applmgmt.log
Forensic Marker: Any HTTP POST request sent to /rest/appliance/access/ssh that changes the SSH daemon status—particularly if initiated by an unexpected internal IP address or followed minutes later by a second request flipping the status back to false—is a strong indicator of unauthorized operational access.
Forensic Check 3: Reviewing vmdird Logs on TCP Port 2012
Inspect the raw directory service audit logs for anomalous connections over the DCE/RPC protocol:
# Review directory daemon logs for anomalous RPC access
grep -i "rpc" /var/log/vmware/vmdir/vmdird.log
Forensic Marker: Look for sudden bursts of unauthenticated connection attempts originating from network zones that have no legitimate business communicating directly with vCenter's internal directory ports. Pay specific attention to connections initiating from jump boxes, internal Linux utility servers, or edge VPN appliances.
Forensic Check 4: Sweeping Physical ESXi Hosts for Injected Keys
If the vCenter server was compromised, assume the threat actor reached downward into managed ESXi hosts.
Security engineers must log into every physical ESXi server and audit the root user's SSH authorization file:
# Inspect authorized SSH keys on bare-metal ESXi hosts
cat /etc/ssh/keys-root/authorized_keys
Forensic Marker: Any unfamiliar cryptographic public key entries in authorized_keys indicate the host has been compromised. The threat actor places keys here to maintain direct access to physical hardware even if the vCenter appliance is subsequently rebooted, rebuilt, or patched.
The Remediation Protocol for Confirmed Breaches
If any of these forensic markers appear, applying Broadcom’s patch is not enough. The enterprise must execute an emergency recovery plan:
- Revoke and Re-issue the SAML Token-Signing Certificate: The threat actor must be assumed to possess the x509 signing key. Re-minting internal certificates invalidates all forged Golden SAML tokens.
- Rotate Machine Accounts and Global Domain Passwords: Reset the passwords for [email protected] and all Active Directory or LDAP service accounts synced with vCenter.
- Re-key Virtual Machine Encryption and KMS: If native vSphere VM encryption is used, re-key the Key Management Server (KMS) relationships to prevent stolen master keys from decrypting offline backups.
- Rebuild the vCenter Server Appliance: The safest operational path is deploying a fresh vCenter virtual appliance from a verified vendor ISO image, re-joining ESXi hosts to the clean instance, and retiring the compromised appliance.
Architectural Hardening: Securing Virtualization Against Future Zero-Days
Security architectures that rely solely on software vendors releasing timely patches will inevitably fail against zero-day campaigns. To protect virtualization assets against future zero-day vulnerabilities, organizations must build layered network defenses that restrict access to administrative interfaces.
HARDENED ZERO-TRUST MANAGEMENT ENCLAVE
[ General Corporate Network ] [ Internet / External Users ]
| |
X (Blocked by Firewall) X (Blocked by Firewall)
| |
v v
========================================================================== MANAGEMENT BOUNDARY
|
+---------------> [ MFA Hardened Privileged Jump Box ]
|
| (Dedicated Out-of-Band Admin VLAN)
v
+-------------------------------+
| Isolated Management Network |
| |
| [ vCenter Server Appliance ] |
| * TCP 443 (Admin UI) |
| * TCP 5480 (VAMI Config) |
| * TCP 2012 (vmdir RPC) |
| | |
+-----------+-------------------+
|
| (Isolated Hypervisor Traffic)
v
+-------------------------------+
| ESXi Hypervisor Hosts |
+-------------------------------+
Complete Network Isolation of Management Ports
vCenter Server appliances must never be exposed to the public internet, nor should they be directly reachable from broad internal employee networks.
Firewalls and internal routing policies must enforce strict access boundaries:
- TCP Port 2012 (vmdir RPC): Must be restricted strictly to localhost and other authorized vCenter servers participating in Enhanced Linked Mode. No general employee workstation, application server, or edge appliance should be permitted to establish a TCP handshake on port 2012.
- TCP Port 5480 (VAMI Management): Should be locked down exclusively to designated systems administration subnets.
- TCP Port 443 (vSphere Web Client): Access should require passage through dedicated, hardened management bastion hosts (Jump Boxes) that enforce hardware token authentication.
Enforcing Microsegmentation
Enterprises must use network microsegmentation to separate the management plane from the data plane.
A compromised virtual machine running inside an ESXi host should have zero network visibility into the vCenter appliance managing that host. By configuring software-defined firewall rules at the hypervisor level (using VMware NSX or equivalent network boundary tools), any network packet moving from a standard virtual machine subnet toward vCenter’s management IP can be dropped and flagged as an attack.
Centralized Off-Appliance Log Shipping
Advanced threat actors alter local log files to cover their tracks. Systems administrators should configure vCenter Server to forward all audit events, authentication attempts, API calls, and system messages continuously to an external, write-once Security Information and Event Management (SIEM) system:
vCenter VAMI Console -> Syslog Configuration -> Export to Remote Collector (TLS/TCP 6514)
By streaming syslog data off the appliance in real time, security teams preserve an immutable audit trail that attackers cannot erase, even if they obtain root administrative authority over vCenter.
Infrastructure on the Frontlines of Geopolitical Conflict
The disclosure of CVE-2026-123890 marks a clear shift in how advanced persistent threat groups target global enterprise networks.
As perimeter defenses improve and endpoint security software monitors guest operating systems, state-sponsored actors are focusing their best exploits lower down the technology stack. By weaponizing zero-day flaws in hypervisors, directory daemons, and virtualization control planes, attackers bypass traditional endpoint monitoring, evade corporate detection engineering, and access organizational data at the foundational layer.
The fact that ARMORY CASTLE operated this VMware vCenter zero-day inside targeted networks for over two months before detection illustrates the reality of modern cyber warfare: sophisticated adversaries often hold silent administrative access to infrastructure long before public advisories appear.
For chief information security officers, infrastructure architects, and systems administrators, the takeaway is straightforward. Modern enterprise defense requires moving beyond assumptions of internal network trust.
Broadcom’s emergency patches must be applied immediately across all affected vCenter and Cloud Foundation environments. But more importantly, the management planes that control the world’s virtualized data centers must be treated as critical targets. They must be isolated behind strict network perimeters, continuously monitored for identity anomalies, and audited under the assumption that an adversary will eventually find a path to their gates.
Reference:
- https://socradar.io/blog/cve-2026-123890-vcenter-0day-armory-castle/
- https://www.rapid7.com/blog/post/etr-cve-2026-123890-critical-authentication-bypass-vmware-vcenter/
- https://www.securityweek.com/newly-patched-vmware-vcenter-flaw-exploited-as-zero-day-by-chinese-apt/
- https://socprime.com/blog/cve-2026-123890-critical-vmware-vcenter-zero-day-exploited-by-china-linked-hackers/
- https://aviatrix.ai/threat-research-center/critical-vmware-vcenter-flaw-cvss-98-lets-network-attackers-bypass-authentication/
- https://thehackernews.com/2026/11/critical-vmware-vcenter-flaw-cvss-98.html
- https://www.vicarius.io/vsociety/posts/understanding-cve-2023-34048-a-zero-day-out-of-bound-write-in-vcenter-server
- https://censys.com/advisory/cve-2024-38813/
- https://kudelskisecurity.com/research/vmware-vcenter-server-out-of-bounds-write-vulnerability
- https://www.securityweek.com/chinese-spies-exploited-vmware-vcenter-server-vulnerability-since-2021/
- https://www.reddit.com/r/sysadmin/comments/1y124yn/vcenter_cve2026123890_was_exploited_as_a_zeroday/
- https://www.scworld.com/brief/broadcom-patches-critical-vmware-vcenter-vulnerability-exploited-by-state-sponsored-hackers
- https://www.sentinelone.com/vulnerability-database/cve-2024-38812/
- https://www.bleepingcomputer.com/news/security/vmware-patches-vcenter-zero-day-exploited-in-attacks-since-september/