China-Linked 'Armory Castle' Exploited VMware vCenter Zero-Day for Two Months Before Patch
A suspected China-nexus cyber-espionage group has been caught exploiting a critical, now-patched authentication bypass in VMware vCenter Server as a zero-day for more than two months before Broadcom shipped a fix — turning a single unauthenticated network request into full administrator control of the virtualization layer that underpins most enterprise data centers.
Google Threat Intelligence Group (GTIG) and Mandiant attributed the activity to a group it tracks as Armory Castle, formerly known as UNC5221 and also referenced elsewhere as Warp Panda. According to the companies' joint technical analysis, the group began weaponizing the flaw in early September 2026 — roughly ten weeks before Broadcom publicly disclosed and patched it on November 19, 2026.
The vulnerability: CVE-2026-123890
The flaw, tracked as CVE-2026-123890, carries a CVSS score of 9.8 (Critical) and lives in the VMware Directory Service, known internally as vmdird. vCenter is the central management console that administrators use to control ESXi hosts, provision virtual machines, and manage the rest of a VMware-based data center — which makes any pre-authentication bypass against it an exceptionally high-value target.
GTIG and Mandiant describe the bug precisely as "an unauthenticated Secure Remote Password (SRP-6a) session key disclosure vulnerability over Distributed Computing Environment/Remote Procedure Calls (DCE/RPC)." In plain terms, an attacker with nothing more than network reachability to a vCenter appliance's RPC interface can recover a session key that SRP-6a is supposed to keep secret, and use it to forge a fully authenticated session.
"Successful exploitation results in a remote authentication bypass which creates an authenticated session established as [email protected]," the researchers wrote. That account is the built-in super-administrator of the vCenter Single Sign-On (SSO) domain — exploitation doesn't yield a limited foothold, it hands the attacker the keys to the entire virtual infrastructure in one step, no credentials, phishing, or social engineering required.
Notably, the researchers flagged that TCP port 2012 — the control-interface RPC endpoint for vCenter SSO — is an integral part of how Armory Castle's exploitation chain actually works. Any organization that exposes this port to broader network segments, or fails to firewall vCenter management interfaces from general user VLANs, should treat that as an immediate audit priority independent of patching.
BONESAW: the exploitation tooling
The attackers didn't rely on manual exploitation. GTIG and Mandiant say Armory Castle used a purpose-built Golang utility they've named BONESAW, "a utility primarily built to facilitate remote exploitation ... which utilizes CVE-2026-123890 to exfiltrate the arbitrary Lightning Memory-Mapped Database (LMDB), hijack the SAML Security Token Service (STS), and create a Single Sign On (SSO) Administrator account."
That description maps to a disturbingly complete attack chain packed into one binary:
- Database exfiltration — BONESAW pulls the LMDB-backed
data.mdbdirectory database straight out of vmdird, which contains the full SSO directory: user objects, group memberships, and critical security material. - SAML/STS hijacking — by obtaining the SAML Security Token Service's signing material, the attackers can mint their own valid SSO tokens, effectively forging trusted authentication assertions across any service that trusts that STS.
- Rogue admin account creation — the tool automates creation of a new SSO administrator account, giving the operators persistent, credentialed access that doesn't depend on re-exploiting the original bug.
Once inside, the reported post-exploitation activity escalated quickly. Mandiant observed that after using a companion tool dubbed BEARBONES to open a listener on TCP port 33992, "Armory Castle moved laterally to the vCenter appliance as the Administrator and created a local administrator account named sps-<REDACTED>@vsphere.local." From there, the actor enabled SSH on the appliance and connected using the newly minted local administrator credentials — a shift from exploiting an application-layer bug to holding durable, interactive shell access on the underlying Linux appliance itself.
The final stage of the observed intrusion moved from the management plane down to the hypervisor layer: the actor enumerated the ESXi file system, mounted Virtual Machine File System (VMFS) volumes directly, and added an SSH public key to the root account's authorized_keys file on ESXi hosts. From there, Armory Castle conducted additional reconnaissance and data theft against files on the ESXi datastore, archiving the stolen material and exfiltrating it via HTTP GET requests routed through the environment's Remote Support appliance — a method likely chosen specifically to blend exfiltration traffic with legitimate vendor-support channels and reduce the odds of detection by network monitoring tuned to flag unusual outbound connections.
In short: one auth-bypass bug in vCenter's directory service gave this actor a path from "no credentials" to "root on every ESXi host and silent exfiltration of VM-level data," without the attacker ever needing a phished password or an insider.
Who's behind it, and why it matters
Armory Castle/UNC5221 is tracked by Google and Mandiant as a China-nexus espionage actor with a track record of targeting edge infrastructure and virtualization platforms — environments that, when compromised, give an intruder visibility and control over everything those systems host, rather than just one endpoint. Exploiting vCenter specifically is a force-multiplier move: compromising the management plane of a virtualization cluster can expose every guest VM running underneath it, including domain controllers, backup systems, and security tooling that themselves run as VMs.
The roughly ten-week gap between first observed exploitation (early September 2026) and public disclosure/patch (November 19, 2026) is also worth sitting with. It means organizations running vulnerable vCenter builds had no vendor signal that anything was wrong for over two months, while a capable, patient actor had administrator-level access. That window is long enough for deep persistence, credential harvesting across adjacent systems, and quiet exfiltration — the kind of dwell time organizations should assume occurred if they were running affected versions and lack strong internal telemetry (SSO audit logs, SSH session logging on appliances, ESXi datastore access logs) going back to early September.
Affected products and fixed versions
Broadcom's advisory, released alongside patches on November 19, 2026, lists the following affected products and remediated versions:
- VMware Cloud Foundation / VMware vSphere Foundation 9.1.x.x — fixed in 9.1.1.0200
- VMware Cloud Foundation / VMware vSphere Foundation 9.0.x.x — fixed in 9.0.2.0200
- VMware vCenter 8.0 — fixed in 8.0 U3m
- VMware vCenter 7.0 — contact Broadcom Support if under an extended support contract (no standard patch path)
- VMware Cloud Foundation (legacy) — fixed in 5.2.5
- VMware Telco Cloud Platform — fixed via vCenter 8.0 U3m (covers TCI 3.0, TCP 3.0, 4.x, 5.0.x, 5.1.x, 5.2)
- VMware Telco Cloud Infrastructure — fixed via vCenter 8.0 U3m (same TCI/TCP coverage as above)
Broadcom's advisory lists no workaround, so network-level mitigation (restricting access to the vCenter management interfaces and the RPC port used by vmdird/SSO) is the only interim control for environments that cannot patch immediately.
What defenders should do now
- Patch immediately. There is no workaround for CVE-2026-123890 — upgrading to the fixed vCenter/VCF/vSphere Foundation build listed above is the only remediation Broadcom offers. Given confirmed in-the-wild exploitation since September, treat this as an emergency change, not a routine maintenance-window item.
- Assume compromise if you were exposed and unpatched. Any vCenter instance that was internet- or broadly network-reachable and running an affected version between September and November 19 should be treated as potentially compromised until proven otherwise.
- Audit SSO and local accounts. Look for unexpected administrator accounts in the
vsphere.localSSO domain, and particularly for local admin accounts following unusual naming patterns (the reported intrusion used ansps-<REDACTED>@vsphere.localaccount). Review SSH enablement history and authorized_keys entries on both the vCenter appliance and ESXi hosts. - Restrict and monitor port 2012 and other vCenter SSO/RPC control interfaces at the network layer; they should never be reachable from general user segments.
- Check Remote Support appliance logs for anomalous outbound HTTP GET traffic, since exfiltration in this campaign was routed through that channel specifically to look like legitimate support activity.
- Pull SAML/STS signing certificates and rotate them if there's any suspicion of compromise, since a hijacked STS can be used to forge trusted authentication assertions well beyond the vCenter environment itself.
CVE-2026-123890 is a textbook example of why authentication-bypass bugs in virtualization management planes deserve emergency-patch treatment: the blast radius isn't one server, it's every workload the hypervisor hosts. Organizations running VMware vCenter, Cloud Foundation, vSphere Foundation, or Telco Cloud products should confirm they are on a fixed build today and begin retroactive log review covering the full exposure window back to early September.