# ARMORY CASTLE (suspected PRC-nexus, formerly UNC5221) Exploits VMware vCenter Zero-Day CVE-2026-123890 and BeyondTrust Zero-Day CVE-2026-102764

> Google Threat Intelligence Group and Mandiant report that suspected PRC-nexus espionage actor ARMORY CASTLE used two zero-days: BeyondTrust Remote Support CVE-2026-102764 for initial access and VMware vCenter Server 7.x/8.x CVE-2026-123890 (unauthenticated SRP-6a session key disclosure over DCE/RPC leading to authentication bypass as administrator@vsphere.local), exploited in early September 2026. The actor deployed custom malware (SLIMSAVAGE, CAGEMATCH, BONESAW, SAWSNAKE, BEARBONES, LAZYBONES) and stole credentials and Active Directory data from guest VMs.

- **Published:** 2026-11-19T14:32:58Z
- **Last reviewed:** 2026-11-19T14:32:58Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-6463
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-6463.png
- **ID:** TL-2026-6463
- **Severity:** CRITICAL (CVSS 9.3)
- **Category:** VULNERABILITY
- **Status:** ACTIVE
- **Actor:** ARMORY CASTLE (China)
- **Detections:** 9 (1 shown free below) · **IOCs:** 26 (3 shown free below)
- **CVEs:** CVE-2026-123890, CVE-2026-102764, CVE-2026-59309

## Description

ARMORY CASTLE (formerly UNC5221) is a suspected PRC-nexus espionage actor that GTIG/Mandiant describe as active since at least October 2023 and targeting a wide range of countries and verticals. In a 2026-10-15 post, GTIG documented exploitation of CVE-2026-102764, an unauthenticated improper origin validation / arbitrary file upload flaw (CWE-290) in BeyondTrust Remote Support affecting versions up to 26.2.3, exploited as a zero-day from late August 2026 (BeyondTrust bulletin BT26-05; SaaS tenants patched 2026-10-09 per GTIG). Post-exploitation on the appliance included the SAWFLY PHP webshell (mangoapi.php, Base32 transport plus RC4, gated on POST parameter _c94mfe), a SUID copy of /bin/sh at /usr/bin/dbus-authd, dumping of site and appliance credential vaults with the native /ns/sbin/bt-secrets utility, sudoers-based persistence through a zz-root-persist post-boot script, extensive log deletion, and the BARECLAW SOCKS5-over-WebSocket tunneler exposed through a modified HAProxy configuration.

In the 2026-11-19 follow-up, GTIG describes pivoting from the Remote Support appliance into VMware infrastructure. After scanning from the appliance for VMware-related ports (9080, 5480, 2012, 443, 389, 22, 445, 3389, 21), the actor dropped the SLIMSAVAGE JSP webshell (ActiveSer.jsp) and later the CAGEMATCH staging webshell (u.jsp) on an internal Linux appliance without EDR, set the SUID bit on /bin/env via gdb to gain root, and staged tooling with curl into /tmp. The BONESAW Go utility exploited CVE-2026-123890 (disclosed by VMware on 2026-11-19; Broadcom advisory 39495; vCenter Server 7.x and 8.x), an unauthenticated SRP-6a session key disclosure over DCE/RPC (vmdird, port 2012) that yields authentication as administrator@vsphere.local. BONESAW exfiltrated /storage/db/vmware-vmdir/data.mdb, extracted the STS SAML signing key, created an SSO administrator and assigned Global Administrator through the vSphere SOAP API. SAWSNAKE, a Go port of the horizon3ai vcenter_saml_login.py, then forged SAML assertions (Golden SAML), bypassing MFA and minting administrator session cookies. The actor additionally ran BEARBONES, a backdoored Dropbear SSH server (banner SSH-2.0-pulsesecure_23345) on ports 53 and 33992 that accepts a hardcoded password and logs plaintext credentials.

On vCenter the actor created an sps-<redacted>@vsphere.local account, enabled SSH via PUT /rest/appliance/access/ssh on VAMI (port 5480), extracted the vCenter master key and PostgreSQL credentials, then disabled SSH. On ESXi they used vpxuser and root SSH, added an SSH key to root authorized_keys, mounted VMFS volumes and used vmfs_browser.py to extract NTDS, SYSTEM, SECURITY and DNS data from guest VMs, archiving to o.tar.gz in a hidden .a directory for exfiltration. The LAZYBONES Go backdoor (/usr/bin/vmtoolscs) used Slack for C2 with AES-256-CTR, Base64 and 150-450 second jitter, supporting cmd, upload, download and sleep, and persisted through a systemd service restarting every 50 minutes; GTIG notes LAZYBONES was previously seen in an early-2025 CVE-2025-22457 campaign. BONESAW carries an embedded build timestamp of 2026-07-13, about 16 days before the distinct vCenter vmdir flaw CVE-2026-59309 was disclosed (2026-07-29), though Go build timestamps can be altered. CVE-2026-59309 is a separate LDAP (389/636) authentication bypass and only overlaps conceptually. Victim sectors and regions are not named in the sources; no file hashes were published for the vCenter-stage malware (full IOCs reside in a GTI collection for registered users).

## MITRE ATT&CK

- T1190 Exploit Public-Facing Application
- T1059.004 Unix Shell
- T1505.003 Web Shell
- T1543.002 Systemd Service
- T1136.001 Local Account
- T1098.004 SSH Authorized Keys
- T1548.001 Setuid and Setgid
- T1070.004 File Deletion
- T1078 Valid Accounts
- T1606.002 SAML Tokens
- T1003.003 NTDS
- T1552 Unsecured Credentials
- T1046 Network Service Discovery
- T1021.004 SSH
- T1560.001 Archive via Utility
- T1090 Proxy
- T1573.001 Symmetric Cryptography
- T1071.001 Web Protocols

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**TL-2026-6463 ARMORY CASTLE BeyondTrust Webshell Access (SAWFLY / SLIMSAVAGE / CAGEMATCH) - Exploitation** (`TL-2026-6463-SPL-001`)

Detects HTTP requests to webshell paths dropped by ARMORY CASTLE after exploiting BeyondTrust Remote Support CVE-2026-102764 and during the vCenter pivot: SAWFLY mangoapi.php, SLIMSAVAGE ActiveSer.jsp, CAGEMATCH u.jsp and the app_84304470 staging PHP file. Any hit on these names against a BeyondTrust appliance or internal Linux appliance indicates post-exploitation webshell use.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1190, T1505.003, T1059.004
- **Data sources:** web:access_combined, web:iis, web:haproxy:http, CommonSecurityLog, webserver
- **Severity:** critical
- **Confidence:** high

### Splunk SPL

```spl
index=web sourcetype IN (access_combined, iis, haproxy:http)
(uri_path="*/mangoapi.php" OR uri_path="*/ActiveSer.jsp" OR uri_path="*/u.jsp" OR uri_path="*app_84304470*")
| eval webshell=case(like(uri_path,"%mangoapi.php"),"SAWFLY PHP webshell (CVE-2026-102764)", like(uri_path,"%ActiveSer.jsp"),"SLIMSAVAGE JSP webshell", like(uri_path,"%/u.jsp"),"CAGEMATCH JSP staging webshell", true(),"app_84304470 staging webshell")
| stats count min(_time) as first_seen max(_time) as last_seen values(http_method) as methods values(status) as statuses dc(src) as distinct_clients values(src) as clients by host, uri_path, webshell
| convert ctime(first_seen) ctime(last_seen)
```

### Microsoft KQL

```kql
CommonSecurityLog
| where TimeGenerated > ago(24h)
| where RequestURL contains 'mangoapi.php' or RequestURL contains 'ActiveSer.jsp' or RequestURL endswith '/u.jsp' or RequestURL contains 'app_84304470'
| extend Webshell = case(RequestURL contains 'mangoapi.php', 'SAWFLY PHP webshell', RequestURL contains 'ActiveSer.jsp', 'SLIMSAVAGE JSP webshell', RequestURL endswith '/u.jsp', 'CAGEMATCH JSP staging webshell', 'app_84304470 staging webshell')
| project TimeGenerated, DeviceVendor, DeviceProduct, SourceIP, DestinationIP, RequestMethod, RequestURL, Webshell
```

### Sigma

```yaml
title: BeyondTrust Appliance Webshell Access - SAWFLY SLIMSAVAGE CAGEMATCH
id: 3b0f6c2e-8d41-4a7e-9a52-6f1c0d7e2a11
status: experimental
description: Detects web requests to webshell paths used by ARMORY CASTLE after exploiting BeyondTrust Remote Support CVE-2026-102764 and during the vCenter pivot (SAWFLY mangoapi.php, SLIMSAVAGE ActiveSer.jsp, CAGEMATCH u.jsp, app_84304470 staging file).
references:
  - https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-beyondtrust
  - https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-vsphere/
author: Threadlinqs Team
date: 2026-11-19
tags:
  - attack.initial_access
  - attack.t1190
  - attack.persistence
  - attack.t1505.003
  - cve.2026.102764
logsource:
  category: webserver
detection:
  selection_path:
    cs-uri-stem|endswith:
      - '/mangoapi.php'
      - '/ActiveSer.jsp'
      - '/u.jsp'
  selection_staging:
    cs-uri-stem|contains: 'app_84304470'
  condition: 1 of selection_*
falsepositives:
  - Legitimate application files that happen to be named u.jsp
  - Vulnerability scanners probing for known webshell names
level: critical
```

**Known false positives:**

- Legitimate application files named u.jsp on internal Java appliances
- Vulnerability scanners probing for known webshell names

**Tuning notes:** Scope to BeyondTrust Remote Support appliances and internal Linux management appliances. Exclude known scanner source IPs. The SAWFLY gate parameter _c94mfe is a POST body field and is not visible in access logs; if the WAF logs request bodies, add it as a corroborating condition. Because the actor deletes on-appliance logs, forward logs off-box.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- TL-2026-6463 ARMORY CASTLE VMware Port Sweep Including vmdird 2012 - vCenter Exploitation Staging (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE ESXi VMFS Guest Disk Credential Extraction (vmfs_browser.py, o.tar.gz) (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE LAZYBONES Slack C2 From vmtoolscs Binary (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE Rogue vCenter SSO Administrator (sps-*@vsphere.local) After CVE-2026-123890 (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE Linux SUID Privilege Escalation, bt-secrets Credential Vault Dump and Staged Tool Execution (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE BARECLAW SOCKS5-over-WebSocket Tunnel via HAProxy [indicator] (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE Linux Appliance Persistence Artifacts (zz-root-persist, dbus-authd, vmtoolscs) (locked: unlock with TL Creds)
- TL-2026-6463 ARMORY CASTLE BEARBONES SSH Backdoor and Obfuscation-Network Node Traffic (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sample indicators (3 of 26)

Defanged, TLP:CLEAR.

- IP address: `103.147.251[.]5`
- File hash: `dc14e215269a93fdbc4c1a80b984c23faf7af0911c6dbb85309796a6f2ae95fb`
- IP address: `103.190.38[.]3`

Full set of 26 indicators unlocks with TL Creds: https://threadlinqs.com/tl-creds

## Sources

- [ARMORY CASTLE Exploits vSphere - Google Cloud Threat Intelligence](https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-vsphere/)
- [ARMORY CASTLE Exploits BeyondTrust - Google Cloud Threat Intelligence](https://cloud.google.com/blog/topics/threat-intelligence/armory-castle-exploits-beyondtrust)
- [BeyondTrust Security Bulletin BT26-05](https://beyondtrust.com/blog/entry/bt26-05)
- [BleepingComputer: BeyondTrust patches zero-day exploited in attacks since August](https://www.bleepingcomputer.com/news/security/beyondtrust-patches-zero-day-exploited-in-attacks-since-august/)
- [NY ITS Advisory 2026-074: Multiple Vulnerabilities in VMware Products (CVE-2026-59309 and others)](https://its.ny.gov/2026-074)
- [Broadcom Security Advisory 38017 (CVE-2026-59309/59310)](https://support.broadcom.com/web/ecx/support-content-notification/-/external/content/SecurityAdvisories/0/38017)
- [NVD: CVE-2026-59309](https://nvd.nist.gov/vuln/detail/CVE-2026-59309)
- [CSO Online: BeyondTrust zero-day tied to Chinese spies puts Remote Support users on alert](https://www.csoonline.com/article/4236421/beyondtrust-zero-day-tied-to-chinese-spies-puts-remote-support-users-on-alert.html)
- [Dark Reading: BeyondTrust Remote Support Zero-Day Continues Trend of Exploited Edge Bugs](https://www.darkreading.com/endpoint-security/beyondtrust-remote-support-exploited-edge-bugs)
- [SOCRadar: CVE-2026-102764 BeyondTrust zero-day](https://socradar.io/blog/cve-2026-102764-beyondtrust-zero-day/)
- [Mandiant VCSA Hardening Tool](https://github.com/mandiant/vcsa-hardening-tool)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-6463
