What happened
Broadcom published VMSA-2026-0009 on 19 November 2026. CVE-2026-123890 lets a network-reachable attacker bypass authentication and establish an session through the vCenter Directory Service. Broadcom assigns a maximum CVSSv3 score of 9.8 and provides no workaround. Broadcom lists fixed vCenter builds 9.1.1.0200, 9.0.2.0200 and 8.0 U3m, while VMware Cloud Foundation 5.x is fixed in 5.2.5 and vCenter 7.0 customers must contact Broadcom Support under extended support.
Mandiant says ARMORY CASTLE exploited CVE-2026-123890 in early September 2026 using BONESAW. Mandiant observed BONESAW extracting /storage/db/vmware-vmdir/data.mdb, obtaining SAML Security Token Service signing material and creating an SSO administrator account. SAWSNAKE then enabled forged SAML assertions and administrative session cookies, allowing MFA bypass and user impersonation. The observed campaign used /rest/appliance/access/ssh to enable SSH and executed /tmp/dbr -R -p 33992 to open the BEARBONES listener. The actor subsequently accessed ESXi, examined VMFS volumes and extracted guest-system material.
Google published the network indicators 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. Published file indicators are ActiveSer.jsp, u.jsp, /tmp/123, /tmp/sss, /tmp/dbr and /usr/bin/vmtoolscs. Published YARA rule names include G_Hunting_Webshell_SLIMSAVAGE_1, G_Hunting_Webshell_CAGEMATCH_1, G_Hunting_Utility_SAWSNAKE_2, G_Hunting_Utility_BONESAW_2, G_Hunting_Backdoor_BEARBONES_1 and G_Hunting_Backdoor_LAZYBONES_1. The public article did not publish file hashes; Google placed additional indicators in a GTI Collection available to registered users.
Attribution posture: Google Threat Intelligence Group attributes the observed activity to the suspected PRC-nexus espionage actor ARMORY CASTLE, formerly UNC5221. Broadcom’s advisory does not mention exploitation, but Mandiant describes multiple investigated incidents and a campaign that combined the vCenter flaw with an earlier BeyondTrust Remote Support compromise. Confirmed activity is therefore stronger than an exploitability inference, while the available evidence does not establish how broadly other actors may possess or use the exploit.
Why this matters now
vCenter is a privileged control plane rather than an ordinary application server. An attacker with administrative SSO access can manage virtual infrastructure, enable appliance services, reach ESXi hosts and inspect guest storage. Mandiant’s evidence shows the observed actor was not limited to proving code execution: it obtained the vCenter database and signing material, created persistent identity paths and reached files inside hosted virtual machines. That placement makes the vulnerability relevant to identity, backup, workload and recovery assumptions simultaneously.
The confirmed exploitation path separates patch urgency from incident closure. Updating the appliance removes the documented authentication bypass, but it does not revoke a forged SAML capability, remove an added SSO administrator, restore stolen master keys or prove that data staged from ESXi was not taken. Organisations that treat a successful patch deployment as the final control risk missing persistence and identity compromise already established inside a highly trusted management environment.
The decision for security leaders
Declare two parallel workstreams: emergency vulnerability removal and an evidence-led compromise assessment. Platform engineering owns controlled patching and availability; incident response owns indicator searches, identity review, evidence preservation and the decision on whether STS or other trust material must be replaced. Neither workstream should close the other automatically.
Treat unsupported vCenter 7.0 as an executive exception, not a delayed patch ticket. Require an isolated management path, a Broadcom-supported disposition and a dated replacement plan. Where forensic visibility is insufficient, leadership should choose between rebuilding the control plane from trusted media and formally accepting that pre-patch compromise cannot be excluded.
Require documented results of the compromise hunt before accepting patch completion as sufficient to close the response.
Evidence of closure
- A reconciled CMDB and network-discovery export shows no unowned vCenter or VMware Cloud Foundation instance.
- Version evidence confirms every supported instance runs a Broadcom-listed fixed build.
- A signed hunt report records indicator, SSO, VAMI, SSH, database and ESXi findings for every instance.
- An approved disposition documents rebuild, trust-material replacement or residual uncertainty for every anomalous system.
The Security.io assessment
The most consequential evidence is not simply that the flaw was exploited, but what the actor did with it. Theft of vCenter database and signing material creates identity persistence that can survive a software update. ESXi access also widens the evidentiary boundary from the appliance to guest disks, domain data, backups and other workloads hosted or administered through the environment.
The published indicators are useful but campaign-specific, so a clean IOC result cannot prove absence of exploitation. Closure requires behavioural and identity evidence: administrator creation history, STS and SSO integrity, VAMI service changes, SSH activation, database access, ESXi root activity and unexplained VMFS reads. Organisations without that telemetry should record the assurance limitation rather than converting uncertainty into a clean finding.
Questions for the morning meeting
- Can infrastructure leadership account for every vCenter and VMware Cloud Foundation instance, including unsupported and management-only systems?
- Has incident response proved that patched systems were not compromised before remediation?
- Are vCenter SSO, VAMI, SSH and ESXi events retained outside the virtualisation control plane?
- Who can authorise isolation when vCenter management availability conflicts with evidence preservation?