Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Screenshot of the Google Threat Intelligence Group and Mandiant blog post titled ARMORY CASTLE Exploits vSphere, dated November 19, 2026, listing authors Jayson Frost, Bryce Greathouse, Dhanesh Kizhakkinan, John Palmisano, Steve Rasch, Daniel Sislo and John Wolfram.
Image: Google Threat Intelligence Group / Mandiant (fair use). The report published on Thursday.

Threat Intel

Mandiant Says China-Nexus ARMORY CASTLE Used a vCenter Zero-Day Inside a Victim Network More Than Two Months Before Broadcom's Patch, Forged SAML Tokens and Pulled Active Directory Files Off ESXi Disks

The flaw, CVE-2026-123890, hands an unauthenticated attacker an administrator session on vCenter. Mandiant says the group reached it from a hacked BeyondTrust appliance, built its exploit tool weeks earlier, ran it against a server already patched for a related July bug, and that Broadcom's advisory published on Thursday does not mention exploitation.

Priya Shah, Threat IntelSingapore7 min read

SINGAPORE - A suspected Chinese espionage group exploited an unknown flaw in VMware vCenter in early September, more than two months before Broadcom released a fix on Thursday, and used it to make itself a single sign-on administrator, forge login tokens and copy Active Directory data straight off the virtual disks of a victim's servers, Google's Mandiant and Threat Intelligence Group said.

The vulnerability, CVE-2026-123890, rated 9.8 out of 10 for severity, sits in the vCenter directory service. Mandiant said it lets a remote attacker with no password obtain an authenticated session as administrator@vsphere.local by exploiting predictable timestamp-based identifiers in vCenter's Secure Remote Password authentication interface over DCE/RPC on TCP port 2012. With that session, an attacker can run commands and pull files, including the directory's data.mdb database.

The researchers attributed the intrusion to ARMORY CASTLE, the group formerly tracked as UNC5221. Mandiant reported to Broadcom the bug it found the group using, SecurityWeek reported, and Broadcom's advisory, which also fixes four flaws in SDDC Manager and VMware Tools, does not mention exploitation in the wild. Charles Carmakal, Mandiant's chief technology officer, said on LinkedIn that the company had investigated multiple incidents involving the group's exploitation of the flaw, according to SecurityWeek. "Organizations should both apply the patches and examine their systems for evidence of compromise," he said.

From a support appliance to the control plane

The intrusion began with a different zero-day. In October, Mandiant described ARMORY CASTLE breaking into BeyondTrust Remote Support appliances through CVE-2026-102764. Thursday's report picks up the story from there: the group scanned the internal network with nmap for ports tied to VMware, including 2012, 5480 and 9080, and moved to internal Linux appliances that had no endpoint detection and response software.

On the first appliance it dropped two web shells, ActiveSer.jsp, which Mandiant calls SLIMSAVAGE and which runs commands when given the password 123, and u.jsp, called CAGEMATCH, which writes uploaded files to /tmp. It became root by using the gdb debugger to set the setuid bit on /bin/env, then pulled a password-protected archive from the hacked BeyondTrust appliance and unpacked the exploit tool.

Mandiant diagram of the ARMORY CASTLE intrusion: reconnaissance from a BeyondTrust Remote Support appliance compromised through CVE-2026-102764, SLIMSAVAGE and CAGEMATCH shells and SUID escalation on a Linux appliance, BONESAW exploiting vCenter through CVE-2026-123890, SAWSNAKE for golden SAML, a BEARBONES SSH listener, LAZYBONES on a second Linux appliance, and lateral movement to vCenter and ESXi for credential and data theft.
Image: Google Threat Intelligence Group / Mandiant (fair use).Mandiant's sequence of events, Figure 1 in the report.

That tool, a Go program Mandiant named BONESAW, exploits CVE-2026-123890 to stream out the vCenter directory database, extract the Security Token Service's private SAML signing key, create a backdoor account and an SSO administrator, and grant Global Administrator rights on the vCenter root folder. A second Go utility, SAWSNAKE, used the stolen key to forge SAML assertions for the administrator account, the technique known as golden SAML, which Mandiant said lets an attacker bypass multifactor authentication and impersonate any user.

Built before a related bug was disclosed

Mandiant said BONESAW was run in early September against a vCenter server that was already patched for CVE-2026-59309, a separate vCenter authentication bypass disclosed on July 29. The recovered binary carries a build timestamp of July 13, 16 days before that disclosure, although the researchers noted that such metadata can be altered. "Mandiant has evidence that ARMORY CASTLE had knowledge of CVE-2026-123890 prior to the disclosure of CVE-2026-59309," the report said. The attackers had named the tool vcenter-db-teamcheck-go.

Screenshot of the Mandiant report listing BONESAW command line arguments and the files it writes to disk, including the exfiltrated data.mdb copy and extracted tenant credential key, certificate and context files, followed by the description of the SAWSNAKE SAML utility.
Image: Google Threat Intelligence Group / Mandiant (fair use).BONESAW's options and on-disk artifacts as documented by Mandiant.

Once inside vCenter, the group created a local administrator account, turned on SSH through a PUT request to the appliance management API at /rest/appliance/access/ssh, took the vCenter master encryption key and PostgreSQL credentials, and switched SSH off again. It then logged in to an ESXi host as vpxuser and root, created a hidden .a directory on a datastore and used a Python tool, vmfs_browser.py, to pull NTDS, SYSTEM and SECURITY files and DNS data from a Windows virtual disk. The archive, o.tar.gz, was copied out and the directory deleted.

To keep access, the group ran BEARBONES, a backdoored Dropbear SSH server that lets in anyone whose password matches a hard-coded hash, logs credentials and identifies itself as SSH-2.0-pulsesecure_23345, and installed LAZYBONES, a Go backdoor at /usr/bin/vmtoolscs that takes encrypted commands from a Slack channel. Mandiant said pivoting on the BEARBONES banner turned up a suspected network of compromised third-party firewalls and routers used to hide the group's traffic.

What defenders can do now

Broadcom's fixes are in vCenter 9.1.1.0200, 9.0.2.0200 and 8.0 U3m, SecurityWeek reported. Mandiant said BONESAW works against vCenter 7.x and 8.x; SecurityWeek noted there is no public patch for 7.0, which reached the end of general support in October 2025.

Mandiant's report gives SOC teams concrete places to look:

  • Connections to TCP 2012 on vCenter from hosts with no reason to talk to its directory service, and internal scans for ports 2012, 5480 and 9080.
  • New SSO or local administrator accounts in vsphere.local, and PUT requests to /rest/appliance/access/ssh, especially SSH being enabled and disabled in a short window.
  • Access to the vCenter master key and PostgreSQL password files, vpxuser or root logins to ESXi from unusual hosts, and hidden directories or tar and Python activity on VMFS volumes.
  • Files named ActiveSer.jsp, u.jsp, /tmp/123, /tmp/sss, /tmp/dbr and /usr/bin/vmtoolscs, the SSH-2.0-pulsesecure_23345 banner, outbound Slack API traffic from Linux appliances, and the 11 network indicators and six YARA rules in the report.
Screenshot of Mandiant's VCSA hardening table showing how inbound iptables DROP policies on management ports and TCP 443, with connection logging to a SIEM, would have blocked ARMORY CASTLE's scans and the use of forged administrative cookies.
Image: Google Threat Intelligence Group / Mandiant (fair use).Part of Mandiant's table mapping vCenter hardening controls to the attack chain.

Beyond patching, Mandiant argued that the firewall in vCenter's own management interface is not enough, because it lacks per-port rules, can be switched off by anyone who reaches that interface, does not log denied connections and ignores outbound traffic. It recommended its VCSA Hardening and Logging Tool, which enforces default-deny iptables rules, egress filtering and remote logging of shell activity at the appliance's Linux layer, with testing and change control before production use.

The researchers said the use of two zero-days and three sets of tailored malware in a single intrusion points to significant resources behind the group, and that BONESAW's build date, more than a month before it was used here, suggests it was developed for broader operations.

About the Author

Priya Shah

Priya Shah covers threat intelligence, intrusion analysis, and adversary tradecraft for SOCtember from Singapore.

Related stories

Threat Intel desk