CriticalCVE-2026-123890 5 Hunt LeadsYARAIOCsRemote AttackerARMORY CASTLEBONESAWSAWSNAKESLIMSAVAGECAGEMATCHBEARBONESLAZYBONESGoogle Threat Intelligence Group / Mandiant

China-Linked Spies Minted Their Own vCenter Admins With a Zero-Day

Mandiant says suspected Chinese espionage group ARMORY CASTLE exploited vCenter authentication bypass CVE-2026-123890 as a zero-day from early September, creating SSO admins, stealing SAML keys and copying Active Directory data out of VM disks. Broadcom's VMSA-2026-0009 fixes it.

Broadcom has patched a critical VMware vCenter authentication bypass, CVE-2026-123890, and Google says a suspected Chinese espionage group was using it as a zero-day back in early September. Mandiant and Google Threat Intelligence Group (GTIG) report that ARMORY CASTLE (formerly UNC5221) reached vCenter from inside a network it had already entered through a BeyondTrust Remote Support zero-day, created its own Single Sign-On administrator, stole the SAML signing key and went on to pull Active Directory database files out of virtual machine disks. BleepingComputer notes Broadcom itself hasn’t confirmed exploitation; Mandiant’s forensic evidence is the basis for that.

For defenders

The bug: CVE-2026-123890 (CVSS 9.8) is in the VMware Directory Service (vmdird). Mandiant says predictable UUIDv1 timestamps in the SRP-6a authentication interface over DCE/RPC (TCP 2012) leak the session key, so an unauthenticated attacker gets a session as administrator@vsphere.local without any password. Enhanced Linked Mode makes no difference. Fixed in vCenter 9.1.1.0200, 9.0.2.0200, 8.0 U3m and 5.2.5 (Cloud Foundation), per VMSA-2026-0009, which also fixes four SDDC Manager and VMware Tools for Windows bugs. No workarounds. Not in CISA KEV yet; public exploit code is reported in 4 GitHub repositories. It’s a separate flaw from July’s CVE-2026-59309: BONESAW ran against a vCenter already patched for that one.

The chain: from the BeyondTrust appliance the group scanned for ports 443, 2012, 5480 and 9080 with nmap, T1046 (Network Service Discovery), then planted JSP web shells on an internal Linux appliance that had no EDR, T1505.003 (Server Software Component: Web Shell), and got root by using gdb to set the SUID bit on /bin/env, T1548.001 (Abuse Elevation Control Mechanism: Setuid and Setgid). A Go tool Mandiant calls BONESAW exploited the bug, copied the directory database (data.mdb), extracted the STS SAML signing key and created an SSO admin, T1210 (Exploitation of Remote Services). SAWSNAKE, a Go port of a public vCenter SAML login tool, forged admin tokens, T1606.002 (Forge Web Credentials: SAML Tokens). The group then switched SSH on through the VAMI API, stole the vCenter master key and PostgreSQL password file, logged into ESXi as root, added an SSH key, T1098.004 (Account Manipulation: SSH Authorized Keys), and used vmfs_browser.py to lift NTDS.dit and registry hives straight out of a domain controller’s disk, T1003.003 (OS Credential Dumping: NTDS).

Backdoors: BEARBONES (backdoored Dropbear SSH with a magic password and banner SSH-2.0-pulsesecure_23345) and LAZYBONES (Go, Slack-based C2, systemd persistence), T1102.002 (Web Service: Bidirectional Communication). Mandiant published six YARA rules; they’re in the Hunt Lead report below, exactly as published.

Network
103.147.251[.]5 ; 103.190.38[.]3 ; 176.9.89[.]66 ; 185.71.15[.]109
118.174.128[.]34 ; 118.174.128[.]36 ; 118.174.128[.]38
220.121.128[.]41 ; 222.108.77[.]44 ; 61.216.56[.]127 ; 81.45.89[.]43
(suspected obfuscation network: likely compromised third-party routers and firewalls)
SSH banner: SSH-2.0-pulsesecure_23345 on high TCP ports

Host
ActiveSer.jsp (SLIMSAVAGE web shell) ; u.jsp (CAGEMATCH upload web shell)
/tmp/123 (BONESAW) ; /tmp/sss (SAWSNAKE) ; /tmp/dbr (BEARBONES) ; /usr/bin/vmtoolscs (LAZYBONES)
/tmp/gdb_file ; /tmp/123.key ; /tmp/123.rar ; /tmp/rar

What it means for you

vCenter is the control room for a company’s virtual servers: whoever holds it can reach nearly every system running on top of it. If you don’t run VMware, there’s nothing for you to do. If your business does, or an IT provider runs it for you, this is a patch-today item, and it’s worth asking that provider directly whether it’s done. The group behind it is patient and quiet, and a patch on its own won’t remove an attacker who got in before it.

Do this now

  1. Install the VMSA-2026-0009 updates on every vCenter and SDDC Manager now.
  2. Make sure TCP 2012, 5480 and SSH on vCenter are reachable only from named admin hosts, not from the whole internal network.
  3. Check vCenter for SSO or local admin accounts nobody recognizes, and ESXi root for SSH keys you didn’t add.
  4. If you also run BeyondTrust Remote Support, treat it as a possible entry point and run the hunts below.

Source article: ARMORY CASTLE Exploits vSphere
Publisher: Google Cloud Threat Intelligence blog, Mandiant and Google Threat Intelligence Group (Jayson Frost, Bryce Greathouse, Dhanesh Kizhakkinan, John Palmisano, Steve Rasch, Daniel Sislo, John Wolfram)
Published: Nov. 19, 2026
URL: https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-vsphere

Executive Summary

ARMORY CASTLE, a suspected PRC-nexus espionage actor that Mandiant and GTIG formerly tracked as UNC5221, has exploited zero-days and n-days in edge and security appliances since 2023, including Citrix NetScaler, Ivanti Connect Secure and, from late August 2026, BeyondTrust Remote Support (CVE-2026-102764). In this campaign the group moved from a compromised Remote Support appliance onto internal Linux appliances without EDR, scanned for VMware management ports, dropped the SLIMSAVAGE command web shell and, about four hours later, the CAGEMATCH upload web shell, and escalated to root by using gdb to set the SUID bit on /bin/env. In early September 2026 it ran BONESAW, a Golang tool that abuses CVE-2026-123890, an unauthenticated SRP session key disclosure in the vCenter Directory Service DCE/RPC interface on TCP 2012, to create an SSO administrator, stream the LMDB directory database and extract the STS SAML signing key; BONESAW’s build timestamp of July 13, 2026 predates the related CVE-2026-59309 disclosure, suggesting it was built for wider use. The group followed with SAWSNAKE for Golden SAML token forgery, the BEARBONES Dropbear-based SSH backdoor (banner SSH-2.0-pulsesecure_23345, also seen on a suspected obfuscation network of compromised third-party firewalls and routers), SSH enablement on vCenter through the VAMI API, theft of the vCenter master key and PostgreSQL credentials, root access to ESXi with an added authorized key, and offline extraction of NTDS.dit and registry hives from guest VM disks with vmfs_browser.py; data was staged back through the Remote Support appliance for exfiltration, and the Slack-controlled LAZYBONES backdoor, previously seen after CVE-2025-22457 Ivanti exploitation, was left on another appliance under systemd. The objective is espionage through long-term control of the virtualization layer. Hunt leads, false positive expectations, Mandiant’s published YARA rules and indicators are detailed below; no file hashes were published openly.

ATT&CK chain: Lateral Movement → Discovery → Persistence → Privilege Escalation → Credential Access → Defense Evasion → Collection → Command and Control → Exfiltration

Hunt Lead Grid

Hunt Lead Level Hunt Lead Name ATT&CK Risk of False Positives
1 (Junior) ARMORY CASTLE Tool, Banner and Obfuscation-Network Sweep T1105 (Ingress Tool Transfer), T1090.003 (Proxy: Multi-hop Proxy) Low. The file paths, YARA rules and the pulsesecure_23345 banner are specific to this toolset; the IP addresses are compromised third-party devices that may also carry legitimate traffic, so date any hit against the September to November 2026 window before acting on it.
1 (Junior) SLIMSAVAGE and CAGEMATCH JSP Web Shells T1505.003 (Server Software Component: Web Shell) Low. Legitimate JSP applications don’t run /bin/sh -c on a request parameter guarded by pwd=123 or write arbitrary uploads to /tmp; the YARA rules anchor on those exact code strings.
2 (Pro) gdb SUID Flip on /bin/env T1548.001 (Abuse Elevation Control Mechanism: Setuid and Setgid), T1059.004 (Command and Scripting Interpreter: Unix Shell) Low. Developers attach gdb to processes, but not on production appliances, not driven by a base64-decoded command file in /tmp, and not ending with /bin/env carrying the SUID bit; the rule anchors on that end state.
2 (Pro) vCenter SSH Enabled Through VAMI and New SSO Admins T1136 (Create Account), T1021.004 (Remote Services: SSH), T1606.002 (Forge Web Credentials: SAML Tokens) Medium. Administrators do occasionally enable SSH on vCenter for support work and create SSO accounts; the rule anchors on SSH being switched on and off within a short window around a newly created admin, or on admin activity with no matching interactive login at the identity provider.
3 (Elite) VMware Port Sweep to ESXi Disk Extraction T1046 (Network Service Discovery), T1210 (Exploitation of Remote Services), T1098.004 (Account Manipulation: SSH Authorized Keys), T1003.003 (OS Credential Dumping: NTDS) Medium. Vulnerability scanners and backup servers touch vCenter and ESXi ports routinely; the correlation anchors on a host outside the approved admin and scanner list connecting to TCP 2012, followed by root SSH to ESXi and access to raw /dev/disks devices.

ARMORY CASTLE Tool, Banner and Obfuscation-Network Sweep

After moving onto an internal Linux appliance, ARMORY CASTLE pulled its tools down from the compromised BeyondTrust appliance with curl, disguising them as /content/r.js: a RAR binary saved as /tmp/rar and an encrypted /tmp/123.rar that unpacked BONESAW as /tmp/123. SAWSNAKE followed as /tmp/sss, BEARBONES as /tmp/dbr (run as /tmp/dbr -R -p 53 and then -p 33992, with host keys read from /tmp/rsa_host_key, /tmp/ecdsa_host_key and /tmp/dss_host_key), an OpenSSH private key as /tmp/123.key, and LAZYBONES as /usr/bin/vmtoolscs under a systemd service that restarts every 50 minutes. BEARBONES identifies itself as SSH-2.0-pulsesecure_23345, and Mandiant found the same banner on high ports of compromised third-party firewalls and routers, eleven of which it lists as a suspected obfuscation network: 103.147.251[.]5, 103.190.38[.]3, 118.174.128[.]34, 118.174.128[.]36, 118.174.128[.]38, 176.9.89[.]66, 185.71.15[.]109, 220.121.128[.]41, 222.108.77[.]44, 61.216.56[.]127 and 81.45.89[.]43. The detection is a retrospective sweep of Linux appliances, vCenter and ESXi for those paths, a scan of ELF files with the four YARA rules below, a search of firewall and NetFlow logs for any traffic with the eleven addresses, and a search of SSH banner or IDS logs for pulsesecure_.

Use EDR where it exists, but expect the important hosts (appliances, vCenter, ESXi) to have none, so plan for file-system collection with a YARA scanner, plus firewall, NetFlow and Zeek SSH logs at the network edge. Sweep back to at least late August 2026, when the BeyondTrust exploitation started, and further if Remote Support logs go back further. Scope first to anything that can talk to vCenter or ESXi, then widen. Don’t block by file name alone: vmtoolscs is chosen to sit beside genuine VMware Tools binaries, so check the path, the systemd unit and the YARA match. The eleven IPs belong to hijacked devices of unrelated owners, so blocking them is reasonable but treat a hit as proof of contact rather than of ownership. Any YARA hit or pulsesecure_23345 banner inside the network goes straight to incident response, with every SSO credential, the SAML signing certificate and ESXi root passwords considered compromised. The banner pivot generalises to the wider obfuscation network the group uses for other intrusions.

rule G_Hunting_Utility_SAWSNAKE_2 {
        meta:
            author = "Google Threat Intelligence Group"
        strings:
            $src1 = "saml_login/mdb.go" ascii
            $src2 = "saml_login/saml.go" ascii
            $src3 = "saml_login/c14n.go" ascii
            $s1 = "[+] Successfully obtained Administrator cookies for %s!" ascii
            $s2 = "[*] Successfully extracted trusted certificate 1" ascii
            $s3 = "[*] Successfully extracted the IdP certificate" ascii
            $s4 = "https://%s/ui/saml/websso/sso" ascii
            $s5 = "effective SSO host for the exchange: %q" ascii
            $s6 = "assertion canonicalized for digest: %d bytes, sha256=%s" ascii
        condition:
            uint32(0) == 0x464c457f and (2 of ($src*) or 4 of ($s*))
    }
rule G_Hunting_Utility_BONESAW_2 {
  meta:
author = "Google Threat Intelligence Group"
  strings:
    $s_mdb_path = "/storage/db/vmware-vmdir/data.mdb" ascii
    $s_srp_prime_n = "AC6BDB41324A9A9BF166DE5E1389582FAF72B6651987EE07FC3192943DB56050" ascii
    $s_uuid_dbcp = "72c780fe-af6b-42ba-9b02-f61f4ccc98ed" ascii nocase
    $c_uuid_dbcp = { fe 80 c7 72 6b af ba 42 9b 02 f6 1f 4c cc 98 ed }
    $s_uuid_srp_verf = "c7e94609-6ab0-4767-8e92-b6485bff1649" ascii nocase
    $c_uuid_srp_verf = { 09 46 e9 c7 b0 6a 67 47 8e 92 b6 48 5b ff 16 49 }

  condition:
    uint32(0) == 0x464C457F
    and $s_mdb_path
    and $s_srp_prime_n
    and ($s_uuid_dbcp or $c_uuid_dbcp)
    and ($s_uuid_srp_verf or $c_uuid_srp_verf)
}
rule G_Hunting_Backdoor_BEARBONES_1 {
      meta:
        author = "Google Threat Intelligence Group"
      strings:
        $c_sha256_len = { BA 0E 00 00 00 4C 89 E7 48 89 EE }
        $s_banner = "SSH-2.0-pulsesecure_" ascii
        $s_rsa_key = "/tmp/rsa_host_key" ascii
        $s_ecdsa_key = "/tmp/ecdsa_host_key" ascii
        $s_dss_key = "/tmp/dss_host_key" ascii

      condition:
        uint32(0) == 0x464C457F
        and ($c_sha256_len or $s_banner)
        and 2 of ($s_*)
    }
rule G_Hunting_Backdoor_LAZYBONES_1 {
    meta:
        author = "Google Threat Intelligence Group"
    strings:
        $fn1 = "main.AesCtfCrypt_fromconf" ascii wide
        $fn2 = "main.AesCtfDecrypt_fromconf" ascii wide
        $fn3 = "main.uploadSlackFile" ascii wide
        $fn4 = "main.getLastMessage" ascii wide
        $fn5 = "main.exec_my" ascii wide
        $api1 = "https://slack.com/api/files.getUploadURLExternal" ascii wide
        $api2 = "https://slack.com/api/files.completeUploadExternal" ascii wide
        $api3 = "https://slack.com/api/conversations.history" ascii wide
        $str1 = "ok,current sleep is:" ascii wide
        $str2 = "get but not function" ascii wide
        $str3 = "readfile fail:" ascii wide
    condition:
        uint32(0) == 0x464c457f and
        (
            3 of ($fn*) or
            (2 of ($api*) and 2 of ($str*))
        )
}

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Command and Control → T1105 (Ingress Tool Transfer), T1090.003 (Proxy: Multi-hop Proxy), T1102.002 (Web Service: Bidirectional Communication) for LAZYBONES; Persistence → T1543.002 (Create or Modify System Process: Systemd Service)

Sources: “ARMORY CASTLE Exploits vSphere”, Mandiant and GTIG, Nov. 19, 2026, https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-vsphere (source article; YARA rules reproduced as published there)

SLIMSAVAGE and CAGEMATCH JSP Web Shells

The first foothold on the internal Linux appliance was ActiveSer.jsp, which Mandiant tracks as SLIMSAVAGE: when the pwd parameter equals 123, it passes the i parameter to ProcessBuilder("/bin/sh","-c",...), returns stdout and stderr inside <pre> and appends EXIT: plus the return code. Roughly four hours later the group added u.jsp, tracked as CAGEMATCH, which parses multipart/form-data uploads, writes the file to /tmp/<filename>, echoes the absolute path and shows a bare upload form when called without a payload. The workflow was upload with CAGEMATCH, then chmod and run with SLIMSAVAGE. The detection logic is any JSP containing those code strings (the YARA rules below), any new .jsp file written under a Java application’s web root on an appliance outside a vendor update, and any Java or Tomcat process spawning /bin/sh -c, especially with an HTTP response that ends in EXIT:.

Scan the web roots of every Java-based Linux appliance with the two rules, and add a file integrity check for new JSP files where the appliance allows it; collect web access logs for requests to ActiveSer.jsp or u.jsp, or any JSP called with a pwd= parameter. Sweep back 90 days at minimum, because web shells outlive the session that planted them and this group came back over weeks. The file names are easy to change, so lean on the YARA strings and the process lineage rather than the names. Deleting the JSP isn’t remediation: by the time it’s found the group has had root, so a hit means rebuilding the appliance from a known-good image and rotating every credential it stored. The Java-spawns-shell pattern generalises to most JSP web shells.

rule G_Hunting_Webshell_SLIMSAVAGE_1 {
      meta:
        author = "Google Threat Intelligence Group"
      strings:
        $s1 = "if(\"123\".equals(request.getParameter(\"pwd\")))" ascii wide
        $s2 = "new ProcessBuilder(\"/bin/sh\",\"-c\",request.getParameter(\"i\"))" ascii wide
        $s3 = "InputStream a = p.getInputStream(), b = p.getErrorStream();" ascii wide
        $s4 = "while((c=a.read(d))!=-1){out.print(new String(d,0,c));}" ascii wide
        $s5 = "catch(Exception e){out.print(\"ERR:\"+e);}" ascii wide

      condition:
        all of ($s*)
    }
rule G_Hunting_Webshell_CAGEMATCH_1 {
      meta:
        author = "Google Threat Intelligence Group"
      strings:
        $s1 = "request.getContentType().split(\"boundary=\")" ascii wide
        $s2 = "String s=new String(d,\"ISO-8859-1\");int fi=s.indexOf(\"filename=\\\"\");" ascii wide
        $s3 = "String fname=new File(s.substring(fs,fe)).getName();" ascii wide
        $s4 = "int so=s.substring(0,ds).getBytes(\"ISO-8859-1\").length;" ascii wide
        $s5 = "File f=new File(\"/tmp\");f.mkdirs();File fout=new File(f,fname);" ascii wide

      condition:
        4 of ($s*)
    }

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Persistence → T1505.003 (Server Software Component: Web Shell); Execution → T1059.004 (Command and Scripting Interpreter: Unix Shell)

Sources: “ARMORY CASTLE Exploits vSphere”, Mandiant and GTIG, Nov. 19, 2026 (source article; YARA rules reproduced as published there)

gdb SUID Flip on /bin/env

To get root on the Linux appliance, the group repeated the trick it used on the BeyondTrust appliance: it ran /bin/sh -c echo IWNobW9kIHUrcyAvYmluL2VudgpxCg==|base64 -d > /tmp/gdb_file, which decodes to a gdb script that runs chmod u+s /bin/env and quits, then attached gdb to a privileged process with /usr/bin/gdb -pid 113757 --command=/tmp/gdb_file. With the SUID bit set on /bin/env, any user can run commands as root through it. The detection logic is three independent signals: an echo ... | base64 -d > into /tmp from a web server or Java parent; gdb started with -pid and --command pointing into /tmp; and, most reliably, /bin/env or /usr/bin/env having the setuid bit at all.

Use auditd or EDR process events with full command lines (watch chmod, ptrace and execve of gdb) and a scheduled file-system check such as find / -perm -4000 compared against a baseline of known SUID binaries. Scope to appliances and servers, where gdb has no business running; on developer workstations, alert only on the SUID end state. The SUID check is near-zero noise and should alert outright, since no distribution ships env setuid. Removing gdb from appliances that don’t need it is a sensible hardening step, but don’t rely on it, since the group can bring its own. Treat a hit as root compromise and escalate. The lead generalises to any attacker who uses a debugger to modify permissions rather than writing an exploit.

Hunt Lead Level: 2 (Pro)

ATT&CK: Privilege Escalation → T1548.001 (Abuse Elevation Control Mechanism: Setuid and Setgid); Execution → T1059.004 (Command and Scripting Interpreter: Unix Shell)

Sources: “ARMORY CASTLE Exploits vSphere”, Mandiant and GTIG, Nov. 19, 2026 (source article)

vCenter SSH Enabled Through VAMI and New SSO Admins

BONESAW’s csa subcommand (flags -h, -u, -p, -lu, -d, -mdb-path, -check-version, -auto-d) creates an SSO administrator, and the tool also creates a backdoor machine account through vmdir and grants Global Administrator on the vCenter root folder through the vSphere SOAP API. It leaves files named anon_<host>_data.mdb and <tenant>-tenantcredential-<N>.key.pem, .key.der, .leaf.pem, .chain.pem and .context.json under temp/extracted_tenant_signer_<host>/ on the machine it runs from, and SAWSNAKE writes idp_key.pem, trusted_cert_2.pem and domain.txt before the group deletes it. On vCenter itself the group created a local administrator named sps-<redacted>@vsphere.local, enabled SSH with a PUT request to /rest/appliance/access/ssh on the VAMI API, logged in over SSH, read the vCenter master symmetric key and the PostgreSQL password file, and disabled SSH again immediately afterwards. The detection logic is a VAMI PUT to /rest/appliance/access/ssh, any new SSO or local admin account or new Global Administrator permission, an SSH login to vCenter shortly after SSH was enabled, and any of those files appearing on a non-vCenter host.

Forward vCenter logs to a SIEM: VAMI and appliance management logs for the SSH toggle, the SSO audit events for account and group changes, vpxd events for permission changes, and /var/log/auth.log or its Photon equivalent for SSH sessions. Sweep back to September 2026 and compare every SSO and local account against a list your administrators confirm. An SSH enable followed within the hour by a disable is a strong signal on its own, since support work tends to leave it on for longer. Because SAWSNAKE forges SAML tokens, an attacker can act as any user without a login at your identity provider; correlate vSphere admin sessions with identity provider sign-ins and treat sessions without one as suspect. Resetting passwords alone won’t evict a Golden SAML actor: if this fires, the STS signing certificate must be replaced along with the SSO admin credentials. Mandiant’s free VCSA Hardening & Logging Tool on GitHub sets up the kernel-level logging this lead needs.

Hunt Lead Level: 2 (Pro)

ATT&CK: Persistence → T1136 (Create Account); Lateral Movement → T1021.004 (Remote Services: SSH); Credential Access → T1606.002 (Forge Web Credentials: SAML Tokens), T1552.004 (Unsecured Credentials: Private Keys), T1552.001 (Unsecured Credentials: Credentials In Files)

Sources: “ARMORY CASTLE Exploits vSphere”, Mandiant and GTIG, Nov. 19, 2026 (source article)

VMware Port Sweep to ESXi Disk Extraction

The intrusion follows a clear sequence. From a compromised appliance the group ran nmap -sS across /24 ranges for TCP 443, 2012, 389, 5480 and 9080 (plus 445, 3389, 22 and 21) with --open, and browsed /sdk and /ui. BONESAW then talked to the rpc_srp_verifier service on TCP 2012 of an internal vCenter (7.x or 8.x). After stealing vCenter’s secrets the group logged into the ESXi Host Client as vpxuser and over SSH as root, enumerated volumes with WinSCP, added a public key to root’s authorized_keys, created a hidden .a folder on a VMFS volume over SCP, unpacked vmfs-pydist.tar.gz and ran python vmfs_browser.py -d "/dev/disks/vml.<id>:1" extract against a domain controller’s -flat.vmdk to copy Windows/NTDS, System32/config/SYSTEM, SECURITY and System32/dns. The results were packed into o.tar.gz, copied off with SCP, and the .a folder was deleted. The correlation to build is: an internal host not on the admin or scanner list probes VMware management ports, then connects to TCP 2012 on vCenter, and within days ESXi records a root SSH login, an authorized_keys change, or any process reading /dev/disks/ or a -flat.vmdk directly.

Join internal NetFlow or firewall logs (sources hitting 2012, 5480 and 9080) with vCenter and ESXi syslog (hostd, auth.log, shell logs) and ESXi file integrity on /etc/ssh/keys-root/authorized_keys. Build an allowlist of hosts that legitimately reach TCP 2012 (other vCenter nodes, PSCs, admin workstations) and alert on anything else; this port has no reason to be open to the general network. Treat any Python process on ESXi working with /dev/disks or VMDK partitions as an incident, since normal operations don’t do this, and look for hidden dot-folders on datastores. Expect gaps: Mandiant notes the VAMI firewall doesn’t log denied connections, so forward iptables logs if you can. Sweep back to late August 2026. Blocking nmap isn’t a meaningful control; restricting management ports to a privileged access network is. A match means assuming domain credentials are stolen (an NTDS copy left the network), which calls for a full Active Directory credential reset, including krbtgt twice. The pattern catches any actor who goes after the hypervisor to reach domain controllers without touching them.

Hunt Lead Level: 3 (Elite)

ATT&CK: Discovery → T1046 (Network Service Discovery); Lateral Movement → T1210 (Exploitation of Remote Services), T1021.004 (Remote Services: SSH); Persistence → T1098.004 (Account Manipulation: SSH Authorized Keys); Credential Access → T1003.003 (OS Credential Dumping: NTDS); Defense Evasion → T1006 (Direct Volume Access), T1070.004 (Indicator Removal: File Deletion); Collection → T1560.001 (Archive Collected Data: Archive via Utility)

Sources: “ARMORY CASTLE Exploits vSphere”, Mandiant and GTIG, Nov. 19, 2026 (source article)

Hash Appendix

ARMORY CASTLE vSphere toolset

Mandiant published no file hashes in the article; its indicators are in a Google Threat Intelligence collection available to registered users only. No hashes were looked up in OTX AlienVault or VirusTotal for this report. Detect with the YARA rules and file paths above.

MD5s

(not available)  # ActiveSer.jsp -> SLIMSAVAGE command web shell on internal Linux appliance
(not available)  # u.jsp -> CAGEMATCH upload web shell, writes to /tmp
(not available)  # /tmp/123 -> BONESAW (vcenter-db-teamcheck-go), CVE-2026-123890 exploitation utility
(not available)  # /tmp/sss -> SAWSNAKE vCenter SAML forging utility (two versions)
(not available)  # /tmp/dbr -> BEARBONES backdoored Dropbear SSH server
(not available)  # /usr/bin/vmtoolscs -> LAZYBONES Slack C2 backdoor

SHA1s

(not available)  # ActiveSer.jsp -> SLIMSAVAGE command web shell on internal Linux appliance
(not available)  # u.jsp -> CAGEMATCH upload web shell, writes to /tmp
(not available)  # /tmp/123 -> BONESAW (vcenter-db-teamcheck-go), CVE-2026-123890 exploitation utility
(not available)  # /tmp/sss -> SAWSNAKE vCenter SAML forging utility (two versions)
(not available)  # /tmp/dbr -> BEARBONES backdoored Dropbear SSH server
(not available)  # /usr/bin/vmtoolscs -> LAZYBONES Slack C2 backdoor

SHA256s

(not available)  # ActiveSer.jsp -> SLIMSAVAGE command web shell on internal Linux appliance
(not available)  # u.jsp -> CAGEMATCH upload web shell, writes to /tmp
(not available)  # /tmp/123 -> BONESAW (vcenter-db-teamcheck-go), CVE-2026-123890 exploitation utility
(not available)  # /tmp/sss -> SAWSNAKE vCenter SAML forging utility (two versions)
(not available)  # /tmp/dbr -> BEARBONES backdoored Dropbear SSH server
(not available)  # /usr/bin/vmtoolscs -> LAZYBONES Slack C2 backdoor

References Appendix

  • “ARMORY CASTLE Exploits vSphere”, Mandiant and Google Threat Intelligence Group, Nov. 19, 2026, https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-vsphere
  • “VMware patches vCenter zero-day exploited in attacks since September”, BleepingComputer, Nov. 19, 2026, https://www.bleepingcomputer.com/news/security/vmware-patches-vcenter-zero-day-exploited-in-attacks-since-september/
  • VMSA-2026-0009, Broadcom (VMware), Nov. 19, 2026, https://github.com/vmware/vcf-security-and-compliance-guidelines/blob/main/security-advisories/vmsa-2026-0009/README.md

CVE at a glance

CVE-2026-123890CriticalKEVPublic exploitBroadcom VMware vCenter
CVSS
9.8 Critical (v3.1)
EPSS
0.8% chance of exploitation in the next 30 days (higher than 58.8% of CVEs)
Exploited in the wild
Yes, on CISA’s KEV list since Nov 20, 2026
Public exploit
Yes, reported in 4 GitHub repositories

Full CVE details

Data from NVD, FIRST EPSS and CISA KEV, refreshed automatically. We report whether public exploit code exists but never link to it: fake proof-of-concepts that infect researchers are common.

More from the Roundup