On 19 November Broadcom published VMSA-2026-0009: five flaws, one of them a 9.8 authentication bypass in vCenter's directory service that lets you in as Administrator with no credentials. So far, another Tuesday. What makes this one different is where the flaw sits. Not in the web portal you publish and watch: in an RPC service listening on port 2012. And the forensic report shows the attacker scanning that exact port, with nmap, from inside the network.
If your management-surface inventory says "vCenter's 443", you are missing a port. And the control that would have stopped this was not a patch, because the patch did not exist: the actor had been exploiting it since early September. It was a firewall rule. One that, moreover, cannot live inside vCenter itself — and that is what this post is about.
What breaks, exactly
CVE-2026-123890 is a CWE-288, authentication bypass using an alternate path, scored 9.8 under CVSS 3.1, in the VMware Directory Service. The official description stops at: an attacker with network access can bypass authentication and gain access as the Administrator user. Mandiant, who reported it to Broadcom, goes further: the flaw is "a cryptographic logic and authorization flaw in VMware's rpc_srp_verifier service over port 2012, exposing a shared session key that can be used to create authenticated sessions as [email protected]".
And with that session, the tool they used — BONESAW, written in Go — does all of this in one pass: streams out the directory's raw LMDB file to parse it at leisure somewhere else, extracts the Security Token Service's private RSA signing key, creates a backdoor machine account via vmdir, forges administrative SAML 2.0 assertion tokens and grants itself Global Administrator on the vCenter root folder through the SOAP API. Put another way: it does not steal a password. It takes the machine that makes passwords.
The real route: nobody rang the doorbell
This is the part worth reading before you decide whether it applies to you. The actor — ARMORY CASTLE, previously tracked as UNC5221, a suspected PRC-nexus group according to Google — did not turn up on vCenter's port 443 from the internet. They got in elsewhere and moved around inside: from the Remote Support appliance they moved to an internal Linux appliance, and there they dropped two JSP webshells (ActiveSer.jsp and u.jsp), a modified Dropbear server at /tmp/dbr listening on ports 53 and 33992 with another vendor's banner faked, and a backdoor at /usr/bin/vmtoolscs taking orders over a Slack channel and relaunching itself every 50 minutes via a systemd service.
From there, they went looking. The commands are transcribed in the report, and they are as unsophisticated as this:
nmap -sS -p 443,2012,389 --open <red>.0/24
nmap -sS -p 443,2012,5480 --open <red>.0/24
nmap -sS -p 443,9080 --open <red>.0/24
nmap -sS -p 9080,2012 --open <red>.0/24
nmap -sS -p 443,2012,389,445,3389,22,21 --open <red>.0/24
A /24 sweep looking for VMware's own ports. No chained exploits, no magic: a scan that any network detection tool spots, if somebody is watching the management network. And when they found 2012 open, they dropped BONESAW. Then, from the ESXi host, they used a VMFS reader to open the virtual machines' flat disks and pull the domain controller's NTDS.dit out of them, plus the SYSTEM and SECURITY registry hives and the DNS configuration. Without powering the VM on, without touching Windows, without any domain agent noticing.
If that last part rings a bell, it is because six days ago we wrote about the other side of the same coin: the hypervisor is the one machine in your estate with no agent on it. Here you can see what that blind spot is for.
The port that is not in your inventory
Port 2012 is not a famous port. Mandiant explains it in one line: "port 2012 is used as the control interface RPC for vCenter Single Sign-On (SSO) and is an integral part of ARMORY CASTLE's remote exploitation of CVE-2026-123890 through BONESAW". In other words, it is the channel vSphere's internal pieces use to talk to each other about identity. Nobody publishes it to the internet on purpose. And almost nobody filters it inside, because it never comes up: when people talk about "securing vCenter", they talk about 443 and 5480.
Here is the full inventory, as the attacker used it. This is not a brochure list of best practices: it is what actually got scanned and used.
| Port | What it is | What it was used for |
|---|---|---|
| 2012 | SSO control interface RPC | Exploiting 123890. The one that matters here. |
| 443 | SOAP API and web portal | Assigning privileges and using the forged session cookies. |
| 389 | LDAP (vmdir) |
It shows up in the sweeps: it gives away where the directory lives. |
| 5480 | VAMI, the appliance management | Enabling and disabling the appliance's SSH. |
| 9080 | Internal vSphere service | One of the three "VMware-specific" ports the sweep looked for. |
| 22 | SSH | Shell once VAMI had enabled it, and theft of the master key and the PostgreSQL password file. |
The firewall that does not count
Here is the point that made us write this post, and it is an argument about design, not about patching. Plenty of people answer all of the above by saying the VCSA already ships with a firewall, configurable from VAMI. Mandiant spends a paragraph taking that apart: "A common misconception in defending the vCenter control plane is that the native VCSA host-based firewall managed via the Virtual Appliance Management Interface (VAMI) provides adequate protection against lateral movement". And the reason should be familiar to anyone who has designed a network: that firewall is administered from the very interface the attacker already controls.
Their recommendation is default-deny iptables rules on the appliance itself, and they justify it with a sentence that holds for any console of any brand: "iptables rules operate independently of the VAMI application, preventing the actor from using the VAMI API (/rest/appliance/access/ssh) to circumvent network security controls". Translated into what we do every day: access control for a console cannot depend on the console. If the rule can be changed with the credentials the attacker just stole, it is not a control: it is a preference.
We go one step further, and this part is our own opinion: rules on that appliance are fine as a last trench, but the filtering that really holds sits outside the machine you are protecting, in the network kit. A vCenter whose 2012 only accepts traffic from other nodes of its own cluster and from one specific administrative jump host does not get exploited, zero-day or no zero-day, and that rule cannot be changed from any appliance API. This is what we build in networks and communications when we segment a management plane: not a list of ports on a spreadsheet, but a separate network with its own way in.
The binary's awkward date
There is a detail in the report that deserves its own paragraph, caveat included. The recovered BONESAW binary carries a build timestamp of 2026-07-13T09:21:24Z, and Mandiant adds the reference that makes it interesting: that is 16 days before CVE-2026-59309 was disclosed. 59309 is the previous authentication bypass in the same VMware Directory Service, also a 9.8, the one that came out in VMSA-2026-0006 in late July, the advisory we wrote about at the time. Same component, second 9.8 of the year, almost four months apart.
The caveat, which the report itself supplies and which bears repeating: "Go build metadata can be modified prior to compilation". That date is not proof, it is an indication. But the pattern is documented in our own archive: in August, five days after that July advisory, the other critical flaw from the same batch was already being exploited in the wild, and we covered it here — "no known exploitation" lasted five days. People who invest work in vmdir do not invest it once.
Patch, yes — and mind the small print
To be clear, because none of the above is an excuse not to update: patch, and patch this week. The fixed versions are 9.1.1.0200 for the 9.1.x branch, 9.0.2.0200 for 9.0.x, 8.0 U3m for vCenter 8.0 and 5.2.5 for Cloud Foundation 5.x. Three bits of small print worth knowing before you open the window: there is no published workaround, the patch cannot be applied with Quick Patch (it is a full update, with its reboot) and vCenter 7.0 has no public patch — Broadcom says to presume it affected and request one through an active extended support contract. Their own FAQ files this as "an emergency change under ITIL methodologies", the formal way of saying do not wait for the next change board.
And now the arithmetic that changes the conversation. If exploitation began in early September and the advisory is dated 19 November, we are talking about more than ten weeks in which there was nothing to apply. No patching policy in the world covers that, and it is nobody's fault: it is the definition of a zero day. What was in your hands during those ten weeks was who could open a socket to 2012.
And while you have the window open, the advisory's other four flaws are the sort that make no headlines and deserve a look anyway: a SQL injection (8.1) and a command injection (7.2) in SDDC Manager, both requiring an already authenticated user, a local privilege escalation to root (6.7) in the same product, and a SAML token signature bypass (6.7) in VMware Tools for Windows. Three of them were reported by a researcher at NATO's cybersecurity centre.
Ten read-only minutes, today
If you run vSphere, this can be checked today and changes nothing. Four checks, in this order:
- Who can reach 2012? From any laptop on the user network, run
nmap -sS -p 443,2012,5480,9080 <vcenter-ip>. If it answers anything other than "filtered", you have your answer and need not read the rest of this list. - SSO accounts you did not create. Review the
vsphere.localdomain for recent administrators. In the documented case one turned up with a service-style name, something likesps-<...>@vsphere.local, which passes perfectly well for legitimate in a long list. - Who turned on the appliance's SSH. In the access logs, a
PUT /rest/appliance/access/sshthat nobody on your team remembers making. It is the move that precedes everything else. - Files where there should be nothing. On the appliance:
/tmp/123,/tmp/sss,/tmp/dbrand/usr/bin/vmtoolscs; plus any access to/storage/db/vmware-vmdir/data.mdbthat does not come from the service itself. Mandiant publishes YARA rules and the IP addresses of the infrastructure used, in case you have something to cross them against.
And one honest warning about the fourth, which is our reading and not a recommendation from the report: if you find something there, the patch does not fix your problem. With the STS signing key out of the building, tokens forged afterwards stay valid until that key is rotated. That part is incident response work, not apt work, and it is why detection on the hypervisor and in the directory belongs in managed EDR/MDR rather than in "the server antivirus has it covered".
When this is not urgent for you
The opposite deserves saying too, and it almost never gets written. If your vCenter already lives on a separate management network, with no route from the user network and entry forced through an administrative jump host, your job is to patch in the next window and get on with your life. No crisis meeting. The rest of this post describes a risk you already paid off when you built the network properly, and that is exactly the argument: the segmentation bill is paid once and collected on days like today.
And the other way round, so as not to sell smoke: if you have three hosts, one vCenter and two people in IT, standing up dedicated admin workstations is over-engineering this quarter. The filtering rule for 2012 and 5480, on the other hand, costs one afternoon and does not cost it again. Start there. What does not work is the most common middle position: management plane on the same flat network as the printers, trusting that nobody scans.
What you can check this morning
Before the close, what we do not know. There is no public victim count for 123890, and that needs saying: Mandiant notes the tool was "likely developed for broader operations", which is a reasonable suspicion, not a number.
What we will take away is the thing you can actually check this morning, and it is a one-line question: who can open a connection to your vCenter's port 2012 today? If the answer is "the whole network", yesterday's patch has not changed that. It has shut a door you did not know was there, in a wall that is still exactly where it was.
Sources (verified at origin): the VMSA-2026-0009 advisory of 19/11/2026, the five CVEs with their scores, the fixed versions (9.1.1.0200, 9.0.2.0200, 8.0 U3m, 5.2.5), the absence of a workaround, the fact that the patch cannot be applied with Quick Patch, the vCenter 7.0 situation and the classification as an "emergency change" under ITIL — Rapid7's analysis of the Broadcom advisory and a breakdown of the five CVEs and their reporters, cross-checked against VMware's official security-advisory repository. The technical detail of the flaw (rpc_srp_verifier over port 2012, the shared session key, the LMDB and STS RSA key theft), the role of port 2012, the nmap commands, the implant paths, the NTDS.dit extraction from ESXi, the 2026-07-13T09:21:24Z build timestamp with its caveat about Go metadata, the quotes on the VAMI firewall and the iptables rules, and the detection artefacts — "ARMORY CASTLE Exploits vSphere", by Mandiant and the Google Threat Intelligence Group. The early-September start of exploitation and the attribution to ARMORY CASTLE (formerly UNC5221) — same report, also covered by BleepingComputer. Translations of the quotes into Spanish and Catalan are ours; the English originals are in the links. The six-port table, the arithmetic of the weeks, the judgement about when this is not urgent and the step to filtering that lives outside the machine you are protecting are ours; the circular dependency of the VAMI firewall is formulated in Mandiant's own report — we are translating it. Photograph ours. We are not a Broadcom/VMware partner or reseller, and we do not sell anybody's licences.
Who can reach your management plane?
Separating the management network from the network people work on is one of the cheapest ways to remove real risk, and you do it once. We build it as part of networks and communications, with the watching done in managed EDR/MDR. If you do not know who can talk to your vCenter today, write to us and we will look at it with you.
Talk to everyWAN