CVE-2026-123890: How ARMORY CASTLE Used a vCenter Zero-Day to Reach ESXi

The CyberSec Guru

CVE-2026-123890 vCenter Zero-Day

If you like this post, then please share it:

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Why your support matters: Zero paywalls: Keep the main content 100% free for learners worldwide.

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

On November 19, 2026, Broadcom published VMSA-2026-0009, a critical advisory covering five vulnerabilities in vCenter, SDDC Manager, and VMware Tools for Windows. The headline bug, CVE-2026-123890, scores 9.8 on CVSS 3.1 and has no workaround. Anyone with network access to an affected vCenter can bypass authentication outright.

Google Threat Intelligence Group and Mandiant released their forensic account alongside the advisory, and it shows why that score understates the problem. A suspected PRC-nexus espionage actor tracked as ARMORY CASTLE, previously designated UNC5221, had been exploiting the flaw since early September 2026 with a purpose-built Go utility named BONESAW. The group never attacked vCenter from the internet. It came in through a compromised BeyondTrust Remote Support appliance, crossed an internal Linux server that no endpoint agent was watching, stole the private key behind vCenter’s SAML identity provider, and finished by reading Active Directory database files directly out of virtual machine disks on ESXi datastores. The intrusion chain combined two zero-days, seven previously undocumented malware families, and one reused backdoor.

I read vCenter in this campaign as an amplifier. A foothold on one help-desk appliance became administrative control of an entire virtualization estate, because the directory service at the heart of vSphere keeps identity, credential, and signing material in one reachable place. Patching closes the door the attackers used. It does not revoke what they carried out through it, and that gap should shape how defenders respond this week.

VMSA-2026-0009: what Broadcom fixed and which vCenter builds are patched

The advisory bundles one critical issue with four lesser ones, and the spread in scores matters for triage.

CVEComponentCVSS 3.1Flaw
CVE-2026-123890vCenter (VMware Directory Service)9.8Authentication bypass, CWE-288
CVE-2026-123891SDDC Manager8.1SQL injection with authenticated read and write access to the database
CVE-2026-123892SDDC Manager7.2Authenticated command injection requiring high privileges
CVE-2026-123893SDDC Manager6.7Local privilege escalation from appliance administrator to root
CVE-2026-123894VMware Tools for Windows6.7SAML token signature bypass

CVE-2026-123890 carries the vector AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H. Arnoldas Radisauskas of NATO’s NCSC reported the three SDDC Manager issues. Noam Pomerantz reported the VMware Tools flaw, which an attacker can exploit only after gaining a local privileged process inside a guest configured with guest alias mapping. Broadcom credits Mandiant for the vCenter finding.

The fixed vCenter builds are 9.1.1.0200 for the 9.1.x branch, 9.0.2.0200 for 9.0.x, and 8.0 U3m for vSphere 8. Cloud Foundation 5.x customers take vCenter 8.0 U3m through the asynchronous patching process in KB 88287 and bring SDDC Manager to 5.2.5. The 9.x SDDC Manager fixes land in 9.1.1.0200 and 9.0.2.0100. Telco Cloud Platform and Telco Cloud Infrastructure remediation is documented in KB463116. VMware Tools for Windows is fixed in 13.1.5.0 for the 13.x line and 12.5.5 for older branches, with 12.4.10 covering 32-bit guests.

vSphere 7.0 is affected, reached end of general support on October 2, 2025, and has no public fix. Extended-support customers have to go through Broadcom, and Broadcom’s supplemental FAQ says to presume 6.5 and 6.7 are affected as well. ESX itself, NSX, VCF Operations, VCF Automation, Workstation, and Fusion are not affected by this advisory. None of these issues is a virtual machine escape.

A few operational details in the FAQ deserve attention before anyone opens a change ticket:

  • The patches are cumulative and include everything from VMSA-2026-0008, so there is no prerequisite sequencing.
  • The vCenter updates are not eligible for vCenter Quick Patch. Expect conventional update paths and longer management-plane downtime, though running workloads continue untouched and only certain key-provider operations pause while vCenter is offline.
  • vCenter 8.0 U3m and 9.0.2.0200 carry build numbers that block a later upgrade to VCF 9.1.1. Organizations mid-upgrade must either jump vCenter to 9.1.1 first or wait for VCF 9.1.2, per KB 67077.
  • Enhanced Linked Mode changes nothing about exposure.
  • For the Tools issue, the FAQ rules out the comfortable shortcuts. Disabling guest operations does not help because the attack path is local to the guest and never goes through vSphere. Stopping or uninstalling VMware Tools is no workaround either.

Guest aliases are VGAuth mappings that bind a vSphere SAML identity to a guest OS account, so Guest Operations can authenticate without a guest password. The vulnerable path involves that mapping machinery, and the fix is updating Tools.

How CVE-2026-123890 works: an SRP authentication bypass in vmdird

The VMware Directory Service, vmdird, is the identity backbone of the vCenter Server Appliance, a Photon Linux VM. It stores SSO users, groups, password hashes, machine and solution accounts, and the certificates behind vCenter’s Security Token Service. Most administrators know it as the LDAP service on TCP 389 and 636. It also exposes a DCE/RPC control interface on TCP 2012 that vCenter components use to talk to single sign-on, and that lesser-known port is where CVE-2026-123890 lives.

📬 Stay Ahead of Cyber Threats

Get the latest cybersecurity news, critical vulnerabilities, threat intelligence, tutorials, and exclusive giveaways delivered straight to your inbox. No spam. Unsubscribe anytime.

One of the RPC services on port 2012, which Mandiant identifies as rpc_srp_verifier, authenticates callers with the Secure Remote Password protocol, version 6a. SRP is an augmented password-authenticated key exchange. The server stores a verifier derived from the password instead of the password itself, both sides exchange ephemeral public values, and each can compute the same premaster secret only if the client genuinely knows the password-derived private value. The shared session key K is hashed out of that secret, and later proof messages convince each side the other holds K. The design’s security rests on the secrecy of the ephemeral private values and the unpredictability of the exchange.

According to GTIG’s analysis, the flaw is a cryptographic logic and authorization error. The SRP interface builds part of the exchange from UUID version 1 identifiers, and UUIDv1 is a timestamp by construction: a 60-bit count of 100-nanosecond intervals since October 1582, plus a clock sequence and node identifier. A timestamp-derived value is the wrong primitive where a protocol needs unguessable input. Its predictability lets an unauthenticated remote attacker recover the shared session key K without knowing any password. The service then treats the attacker’s session as a completed password proof, and the session it grants is [email protected], the top of the vSphere identity tree.

Broadcom and Mandiant have not published the precise internal role the UUID plays, and no public proof-of-concept existed as of November 20. What is documented is the primitive, the port, and the outcome.

How CVE-2026-123890 differs from CVE-2026-59309

CVE-2026-59309, fixed in VMSA-2026-0006.1 on July 29, 2026, was also a vmdir authentication bypass that produced rogue SSO accounts. It worked through malformed authorization tokens in crafted LDAP packets on ports 389 and 636. The November flaw shares the outcome, including recovery of K without the password, but travels the DCE/RPC path instead.

Mandiant’s logs show BONESAW running against a vCenter that had already been patched for the July issue. Organizations that finished that patch in August were still exposed through a different interface of the same daemon. The sibling directory-traversal bug from the July advisory, CVE-2026-59310, was exploited quickly enough that CISA added it to the KEV catalog on August 18. CVE-2026-123890 itself was not yet in KEV as of November 20, which says more about catalog latency than about risk.

BONESAW YARA constants and the port 2012 exposure question

GTIG’s published YARA rule for BONESAW doubles as a specification of the protocol surface. The binary embeds the string /storage/db/vmware-vmdir/data.mdb, the hexadecimal prefix AC6BDB41324A9A9BF166DE5E1389582F... that matches the opening bytes of the standard 2048-bit SRP group prime in RFC 5054, and two DCE/RPC interface UUIDs in both ASCII and little-endian binary form. 72c780fe-af6b-42ba-9b02-f61f4ccc98ed is the vmdirdbcp database-copy interface, and c7e94609-6ab0-4767-8e92-b6485bff1649 is the SRP verifier interface.

Those constants make durable hunting anchors on disk. On the wire the practical question is simpler: which sources can open a TCP connection to port 2012? In a well-segmented estate the answer is vCenter nodes and nothing else. ARMORY CASTLE reached it from a compromised internal server, which is the scenario segmentation exists to prevent.

Stage one: the BeyondTrust Remote Support zero-day (CVE-2026-102764)

The entry point was BeyondTrust Remote Support, the remote-access platform formerly known as Bomgar that IT and help-desk teams use to reach endpoints across the enterprise. CVE-2026-102764, disclosed October 15, 2026 and scored 9.3 under CVSS v4, is an improper-origin-validation flaw (CWE-290) in an internal messaging component. The appliance did not adequately verify where internal messages came from, which under the right conditions let an unauthenticated remote attacker perform file operations.

All Remote Support versions up to 26.2.3 were affected. The BT26-05-RS patch covers 24.3.1 through 26.2.3, version 26.3.1 incorporates the fix, and BeyondTrust patched its SaaS fleet on October 9. Mandiant’s October report dates ARMORY CASTLE’s zero-day exploitation to late August, and CISA added the CVE to KEV on October 16 with a three-day federal remediation deadline.

The post-exploitation work on that appliance previews everything that followed:

  • After initial POST requests to a staging file, the actor installed SAWFLY, an obfuscated PHP webshell named mangoapi.php, in the site’s Laravel web root. SAWFLY answers only to requests carrying the parameter _c94mfe, and anything else gets a 404, which keeps scanners and crawlers blind. Supplied payloads are Base32-decoded, decrypted with RC4 under a hardcoded 32-byte key, and executed through @eval(), with results encrypted back along the same channel.
  • Privilege escalation came from copying /bin/sh to /usr/bin/dbus-authd, chowning it to root, and setting the SUID bit.
  • Persistence came from a startup script, zz-root-persist, planted in the appliance’s post_boot.d directory. It appends a NOPASSWD: ALL sudoers entry for the site service account to /ns/config/sudoers.template and forces a rebuild of the live sudoers file with the proprietary bt-sysconfig utility, so the grant survives reboots and configuration resets that would erase a direct edit.
  • The actor tried to dump the appliance’s credential vault with the native bt-secrets CLI, deleted auth and SSL logs throughout, and raised a compromised administrator’s Jump Item and team permissions to Administrator, which turned the product’s own remote-access fabric into a lateral movement mechanism.
  • For external reachability it deployed BARECLAW, a C tunneler that provides SOCKS5 over WebSocket through an embedded CivetWeb server, and edited /ns/config/haproxy.conf so the appliance’s own reverse proxy would route internet traffic to it.

A remote-support appliance is a credential broker with sessions into everything, by design. Owning one is close to owning the network.

Lateral movement: two webshells and a debugger on unmonitored Linux appliances

From the BeyondTrust box, ARMORY CASTLE ran nmap sweeps across internal ranges for ports 443, 2012, 389, 5480, and 9080, and probed the /sdk and /ui paths that mark vSphere endpoints. Port 2012 is a fingerprint of intent: it has no purpose in generic reconnaissance and every purpose in this exploit chain. The actor then moved to internal Linux appliances that lacked EDR coverage, where its tooling could stage without host-level telemetry.

Foothold tooling arrived as a pair of JSP webshells that split the work between them. SLIMSAVAGE, dropped as ActiveSer.jsp, executes commands gated on the hardcoded secret pwd=123. It runs them through ProcessBuilder("/bin/sh", "-c", ...), streams back both stdout and stderr, and appends the process exit code. About four hours later came CAGEMATCH, u.jsp, a dedicated upload handler that parses the multipart/form-data boundary, writes raw file bytes to /tmp/<filename>, and echoes the absolute path. With one shell to upload and the other to chmod and execute, the actor had a complete staging workflow over plain HTTP.

Privilege escalation reused the SUID trick from the BeyondTrust appliance, this time delivered through the GNU Debugger. The actor wrote a base64-encoded command file and attached gdb to a running process:

/bin/sh -c echo IWNobW9kIHUrcyAvYmluL2VudgpxCg==|base64 -d > /tmp/gdb_file
/usr/bin/gdb -pid 113757 --command=/tmp/gdb_file

The blob decodes to two lines, !chmod u+s /bin/env and q. The leading exclamation mark is gdb’s shell-escape command, so the chmod runs in the debugger’s security context instead of as a child of a web server, and then the debugger quits. Once /bin/env carries a root-owned SUID bit, any local user can run programs with root’s effective user ID through it. This is a GTFOBins-style abuse that needs no kernel exploit and leaves one easily checked filesystem artifact: /bin/env should never show -rwsr-xr-x.

With root established, the actor pulled a RAR binary and a password-protected archive from the BeyondTrust appliance using curl -k, both disguised with a .js extension under a /content/ path, and extracted the archive to /tmp/123. That file was BONESAW.

BONESAW: stealing vCenter’s data.mdb identity database

BONESAW is a Golang exploitation utility. Its internal name, vcenter-db-teamcheck-go, and its embedded build path, E:/CodeFiles/VulnScan/Vcenter/go_teamcheck/, point to a Windows development environment. The recovered sample’s Go build timestamp reads 2026-07-13T09:21:24Z, sixteen days before the July disclosure of CVE-2026-59309 and roughly seven weeks before its observed use. GTIG cautions that Go build metadata can be tampered with before compilation. Taken together with the deployment timeline, the timestamp suggests a capability developed in advance for broader operations and not improvised for one victim. GTIG assesses that the tool works against vCenter 7.x and 8.x appliances.

Its primary csa subcommand, described in the tool’s own help text as “Create SSO admin with minimal parameters,” chains the whole compromise. Its flags cover the target’s vim25 SOAP endpoint, the username and password for the new SSO account, a -lu flag naming the existing local admin UPN used for the vmdird bridge authentication (defaulting to [email protected]), tenant discovery via -auto-d, a version probe via -check-version, and -mdb-path to reuse an already-exfiltrated database instead of downloading it again.

Execution proceeds from the SRP session-key disclosure on port 2012 to an authenticated administrator session. BONESAW then abuses the vmdirdbcp RPC interface to stream out /storage/db/vmware-vmdir/data.mdb, the Lightning Memory-Mapped Database file that holds the entire vmdir identity store. LMDB’s design, a memory-mapped copy-on-write B+tree in a single file, is excellent for embedded performance and catastrophic for blast radius. One file read yields users, password hashes, machine accounts, tenant configuration, and the RSA private key the Security Token Service uses to sign SAML assertions.

BONESAW writes the exfiltrated database to disk as anon_<host>_data.mdb and parses tenant signing credentials into <tenant>-tenantcredential-<N> files in PEM and DER form, with the leaf certificate, chain, and a JSON metadata blob alongside. It then creates a backdoor machine account through vmdir and assigns Global Administrator privileges on the vCenter root folder through the vSphere SOAP API.

None of this targeting was new, and that is the part that should sting. Horizon3’s Zach Hanley documented in October 2021 that the IdP signing certificate and key sit in cleartext inside data.mdb, that vCenter file-based backups contain the same database inside lotus_backup.tar.gz, and that possession of the key permits forging a login for any user, including the built-in Administrator. The associated tooling has been public on GitHub ever since, and CVE-2020-3952 had already demonstrated unauthenticated vmdir data exposure years earlier. ARMORY CASTLE already knew data.mdb held the crown jewels. What it added was a reliable, unauthenticated, network-only way to take them.

Technical reconstruction of ARMORY CASTLE's attack chain, showing how CVE-2026-123890 enabled unauthorized vCenter authentication, vmdir database theft, STS signing-key compromise, forged SAML sessions, ESXi management-plane access, and offline extraction of Active Directory credentials from virtual disks
Technical reconstruction of ARMORY CASTLE’s attack chain, showing how CVE-2026-123890 enabled unauthorized vCenter authentication, vmdir database theft, STS signing-key compromise, forged SAML sessions, ESXi management-plane access, and offline extraction of Active Directory credentials from virtual disks

SAWSNAKE and Golden SAML in vSphere

The follow-on utility, /tmp/sss, is tracked as SAWSNAKE. It is a Go port of the public vcenter_saml_login.py tool, and it converts the stolen database into durable access. In extract mode it parses data.mdb and writes idp_key.pem, trusted_cert_2.pem, and domain.txt to an export directory. In direct-cert mode it loads those files, forges a SAML 2.0 assertion for the Administrator account, signs it with the genuine STS private key, and completes the web SSO exchange at the target’s /ui/saml/websso/sso endpoint to collect administrative session cookies.

A second, apparently newer sample added verbose debug logging across the TLS, HTTP, SAML initiation, XML signing, and login routines, plus redirect handling for up to ten HTTP 302 hops to survive multi-hop authentication flows. After use, the actor deleted the tool and its output with rm.

This is Golden SAML applied to vSphere, and its properties explain why a patch ticket cannot close the incident:

  • The forged assertion is cryptographically indistinguishable from one the real identity provider issued, because it is signed with the real identity provider’s key.
  • MFA gates the issuance of an assertion, and the attacker skips issuance by minting their own.
  • Password resets do nothing, because no password is involved.
  • The stolen key keeps producing valid administrator sessions until the STS signing certificate itself is replaced.

Replacing that certificate has a blast radius of its own. Solution users, registered extensions, federated identity providers, and dependent VCF components all trust the old certificate and have to be re-trusted deliberately. Any organization that finds evidence of data.mdb exfiltration should treat signing-key replacement as mandatory incident response and plan the dependency work that comes with it.

BEARBONES and LAZYBONES: backdoors built to blend in

Between exploitation runs, the actor maintained its own access infrastructure on the compromised Linux appliances.

BEARBONES, /tmp/dbr, is a C backdoor built from the open-source Dropbear SSH server. The actor modified its password authentication handler, svr_auth_password, so that SSH_MSG_USERAUTH_REQUEST packets skip /etc/shadow validation when the SHA-256 digest of the first fourteen bytes of the client-supplied password matches a hardcoded hash. The same handler logs every plaintext password it sees, which quietly harvests legitimate administrators who use the same listener. Host keys load from /tmp instead of /etc/dropbear, and the SSH identification banner was changed to SSH-2.0-pulsesecure_23345. The actor ran it first on TCP 53, then on 33992.

Pivoting on that banner, Mandiant mapped a suspected obfuscation network of compromised third-party firewalls and routers exposing the same string on non-standard high ports, with eleven relay IPs published in the accompanying GTI collection. The banner also marks lineage. This is the actor that exploited the Pulse Connect Secure buffer overflow CVE-2025-22457 in early 2025, the intrusion where LAZYBONES first appeared.

LAZYBONES was deployed on a second Linux appliance as /usr/bin/vmtoolscs, a name chosen to look like a VMware Tools component, with a systemd service configured to restart every fifty minutes. Written in Go, it uses Slack as its command-and-control channel, polling conversations.history for the latest message in a designated channel and moving files through Slack’s external upload API endpoints. This variant encrypts C2 messages with AES-256-CTR before Base64 encoding, beacons at a jittered interval of 150 to 450 seconds, and supports four commands: cmd, upload, download, and sleep.

To a network monitor, the traffic is HTTPS to a sanctioned SaaS platform. That is why egress policy for infrastructure appliances should be an explicit allowlist, and should never rest on the assumption that servers do not browse the web.

From vCenter to ESXi: extracting ntds.dit from VMDK files

With administrative sessions established, the actor created a local account named sps-<redacted>@vsphere.local. It then enabled SSH on the vCenter appliance through a PUT request to the VAMI REST endpoint /rest/appliance/access/ssh, a technique Mandiant has also documented in BRICKSTORM intrusions attributed to UNC6201. Over SSH it stole the appliance’s internal master crypto key file and the PostgreSQL password file, then immediately disabled SSH through the same API, which compressed the detectable window to minutes.

Next came authentication to the ESXi Host Client as vpxuser and SSH to the hosts as root. The vpxuser account is the identity vCenter provisions on every managed host. It holds the Admin role, and vCenter rotates its password automatically and stores it in its own database. An interactive Host Client login as that account is a strong anomaly signal because it should never be a human. GTIG does not detail how root credentials were derived, but the sequence, master key and database credentials out and host-level authentication in, describes an operator working down the management plane’s own trust gradient.

On the hosts, the actor enumerated disks and VMFS volumes with WinSCP, appended its own public key to root’s authorized_keys, created a hidden staging directory .a on a datastore, and pulled a VMware-specific toolset onto the host with tar xzf vmfs-pydist.tar.gz. The centerpiece was vmfs_browser.py, a Python tool that reads guest storage from under the running operating system by addressing raw devices and parsing partitions inside flat VMDK files:

python vmfs_browser.py -d "/dev/disks/vml.<REDACTED>:1" extract "<REDACTED>-flat.vmdk:part4/Windows/NTDS" ./NTDS
python vmfs_browser.py -d "/dev/disks/vml.<REDACTED>:1" extract "<REDACTED>-flat.vmdk:part4/Windows/System32/config/SYSTEM" ./SYSTEM

The recorded invocations extract the NTDS directory containing ntds.dit, the SYSTEM and SECURITY registry hives, and the DNS server’s zone directory from a Windows guest. This is the campaign’s quietest and most consequential trick. Nothing inside the guest ever executes. There is no logon event on the domain controller, no file-lock contention on the live database, and no EDR telemetry, because to a root user on the hypervisor the disk is just bytes on a datastore. The SYSTEM hive supplies the boot key needed to decrypt ntds.dit, so the pair amounts to wholesale domain credential compromise without touching the domain controller.

The loot was archived as o.tar.gz, copied back with SCP in the protocol’s internal source mode (scp -r -p -d -f), downloaded over HTTP GET to the BeyondTrust appliance, and likely exfiltrated from there. The actor then deleted the .a directory. The matching private key, /tmp/123.key, stayed on the Linux appliance for passwordless re-entry.

On boundaries: this was not a VM escape, and Broadcom’s FAQ states that none of the advisory’s issues permits guest-to-host code execution. The actor reached the hypervisor the way the architecture permits, through the management plane, using stolen trust.

Why the vCenter appliance firewall failed

GTIG’s hardening commentary takes aim at a common assumption: that the VCSA’s built-in firewall, managed through the VAMI on port 5480, protects against lateral movement. It has four structural weaknesses, and this campaign walked through them in order.

  • Its GUI grants trust per IP and not per port, so a backup server allowed to reach the vSphere API on 443 is typically trusted for SSH on 22 and the VAMI itself.
  • Its management plane is circular. Whoever holds VAMI access can switch the firewall off, and this actor held VAMI access.
  • It produces no remote log of denied connections, so the SOC is blind to the port scans that opened the operation.
  • It filters inbound traffic only, which leaves implanted malware with unmanaged egress.

Mandiant’s answer is the VCSA Hardening & Logging Tool, published alongside its vSphere and BRICKSTORM defender’s guide. It enforces policy at the Photon kernel’s iptables layer instead of through the appliance application. Mapped against this intrusion, the control set cuts the chain at several independent points:

  • A default-deny inbound policy with per-port allowlists turns the nmap sweeps into silent drops, and a log-bridge service forwards those drops to the SIEM, exposing reconnaissance before exploitation.
  • Restricting TCP 2012 to trusted sources removes the RPC path to vmdird entirely. That contains the vulnerability without fixing it.
  • A TRUSTED_WEB_IPS allowlist on 443 makes SAWSNAKE’s cryptographically valid cookies useless from an untrusted source, because network policy does not care what the identity layer believes.
  • Dedicated SSH and VAMI rules block both the API call that enables SSH and any later shell login.
  • An egress DROP policy stops SCP staging and Slack-based C2.

Kernel-enforced rules also survive the compromise of vSphere or VAMI credentials, which decouples network security from application-layer identity. Remote command auditing survives the .bash_history deletion the actor practiced. None of this replaces the patch, and iptables changes on a production appliance need testing and formal change control, as GTIG itself stresses. The design principle still generalizes: on the machines that hold the keys, decide network trust somewhere the keys cannot reach.

How to patch vCenter and hunt for ARMORY CASTLE

The response to VMSA-2026-0009 has to run as two workstreams, because the fixed builds and the intrusion’s persistence mechanisms are separate problems.

Patch vCenter, SDDC Manager, and VMware Tools

Broadcom classes the vCenter update as an emergency change, and Rapid7’s emergent threat response echoes that. The fixed builds are 9.1.1.0200, 9.0.2.0200, and 8.0 U3m, with the VCF 5.x asynchronous path and the Back-in-Time caveat described above.

Confirm current builds with vpxd -v on the appliance or $global:DefaultVIServer.Build in PowerCLI, and check the interoperability matrix against your ESX and vSAN versions. Broadcom renamed the hypervisor from ESXi to ESX starting with the 9.0 releases, even where patch identifiers still say otherwise.

Until patching completes, restricting TCP 2012 to vCenter nodes is the highest-value interim control, and it costs almost nothing operationally because nothing else legitimately needs that port.

Hunting checklist: vCenter appliance, ESXi, and network

Assume compromise until evidence says otherwise. The lookback window starts no later than September 1, 2026, and Remote Support exposure stretches back to late August.

Identity and SSO:

  • Audit the SSO domain for accounts nobody can account for, particularly sps- prefixed administrators, and review permission grants on the vCenter root folder.
  • vCenter has written SSO principal-management events to /var/log/audit/sso-events/audit_events.log since 6.7 Update 2. Forward that file and read it.

vCenter appliance:

  • Check /tmp for the documented tool paths: 123, sss, dbr, 123.key, gdb_file, rar, and 123.rar.
  • Look for ActiveSer.jsp and u.jsp.
  • Verify /bin/env has no SUID bit.
  • Enumerate systemd units for /usr/bin/vmtoolscs.
  • Reconcile every VAMI SSH enablement against a change ticket. The enable, extract, disable pattern shows up only as a pair of API events minutes apart.

ESXi hosts:

  • Review root’s authorized_keys, root SSH logins outside maintenance, and interactive vpxuser logins.
  • Look for hidden directories on VMFS volumes and for vmfs_browser.py, vmfs-pydist.tar.gz, or o.tar.gz.

Network:

  • Sweep for the SSH-2.0-pulsesecure_ banner on high ports.
  • Review connections to GTIG’s published relay addresses.
  • Treat Slack API traffic originating from server appliances as guilty until proven innocent.

If you find evidence of compromise

Order of operations matters here:

  1. Preserve logs, /tmp, systemd units, and appliance snapshots before changing anything, since this actor deletes logs as a habit.
  2. Rebuild instead of cleaning wherever compromise is confirmed or cannot be excluded.
  3. Replace the STS signing certificate and reset SSO credentials. Forged assertions otherwise outlive every other remediation step.
  4. Rotate ESXi root passwords and re-audit host trust.
  5. If domain controllers ran on affected hosts, treat Active Directory as compromised: reset the KRBTGT account password twice, rotate privileged and service credentials, and review authentication activity since the exposure window.
  6. Add anything the BeyondTrust appliance stored or brokered, including credentials typed into support sessions, to the rotation scope.

As of November 20, no public exploit for CVE-2026-123890 exists. The root cause is now published and the vulnerable interface is identified, so patch diffing will do the rest of the attacker’s homework. The time during which this remains a private capability is a courtesy and gives defenders no protection.

CVE-2026-123890 FAQ

What is CVE-2026-123890?

CVE-2026-123890 is a critical authentication bypass (CVSS 3.1 score 9.8, CWE-288) in the VMware Directory Service of VMware vCenter. An attacker with network access to an affected vCenter can obtain an [email protected] session without a password. Broadcom disclosed it on November 19, 2026 in VMSA-2026-0009 and lists no workaround.

Which vCenter versions fix CVE-2026-123890?

The fixed builds are vCenter 9.1.1.0200 for 9.1.x, 9.0.2.0200 for 9.0.x, and 8.0 U3m for vSphere 8. vSphere 7.0 is affected and has no public fix, and Broadcom says to presume 6.5 and 6.7 are affected too.

Who is ARMORY CASTLE?

ARMORY CASTLE is a suspected PRC-nexus espionage actor previously tracked as UNC5221. Google Threat Intelligence Group and Mandiant report that it exploited CVE-2026-123890 as a zero-day from early September 2026 using the BONESAW utility, after gaining initial access through BeyondTrust Remote Support zero-day CVE-2026-102764.

Is CVE-2026-123890 a VM escape?

No. Broadcom’s FAQ states that none of the issues in VMSA-2026-0009 permits guest-to-host code execution. ARMORY CASTLE reached ESXi hosts through the vCenter management plane and read virtual disk contents from the datastore as root.

Does patching vCenter remove ARMORY CASTLE’s access?

No. If the attacker exfiltrated data.mdb, the stolen STS signing key keeps working for forged SAML assertions until the signing certificate is replaced. Patching alone does not revoke that access.

Buy me A Coffee!

Support The CyberSec Guru’s Mission

🔐 Fuel the cybersecurity crusade by buying me a coffee! Your contribution powers free tutorials, hands-on labs, and security resources.

Why your support matters:
  • Writeup Access: Get complete writeup access within 12 hours
  • Zero paywalls: Keep the main content 100% free for learners worldwide

Perks for one-time supporters:
☕️ $5: Shoutout in Buy Me a Coffee
🛡️ $8: Fast-track Access to Live Webinars
💻 $10: Vote on future tutorial topics + exclusive AMA access

“Your coffee keeps the servers running and the knowledge flowing in our fight against cybercrime.”☕ Support My Work

Buy Me a Coffee Button

If you like this post, then please share it:

Exploits

Discover more from The CyberSec Guru

Subscribe to get the latest posts sent to your email!

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.