Why the vCenter Appliance Firewall Is Not Enough
The vCenter appliance firewall cannot defend the control plane: an attacker at the management interface turns it off. Isolate vCenter upstream instead.
The vCenter appliance firewall is not enough because the built-in firewall managed through the appliance management interface was designed for connectivity, not defense, and an attacker who reaches the management interface can simply switch it off. The ARMORY CASTLE campaign behind CVE-2026-123890 showed why this matters: once inside an internal network, the group reached vCenter's management ports directly and used the management interface to enable SSH. A control plane that holds authority over every virtual machine needs network isolation that does not depend on the appliance defending itself.
Why is the built-in vCenter appliance firewall not enough?
The VMware vCenter Server Appliance (VCSA) ships with a host firewall you configure through the VAMI, its web-based appliance management interface on TCP port 5480. Mandiant's analysis of the CVE-2026-123890 intrusions lists four reasons that default firewall does not stop a determined attacker.
First, it lacks port-specific granularity. To let one internal server reach the vSphere API on port 443, administrators often end up trusting that server's IP address for all ports. If that server is later compromised, the attacker inherits a clear path to SSH on port 22 and the VAMI on port 5480.
Second, it has a circular dependency. The firewall is managed through the VAMI, so an attacker who gains VAMI access can deactivate the firewall with the same access. Mandiant saw the group use the VAMI API to enable SSH during the intrusion.
Third, it has forensic gaps. The VAMI firewall manages connectivity and does not forward logs for denied connections or shell activity to a remote collector, so a security team sees nothing when an attacker scans or probes the appliance.
Fourth, it is inbound-only by default. Outbound traffic is unmanaged, which lets malware on a compromised appliance reach external command-and-control infrastructure freely.
What does a stronger control plane boundary look like?
Treat vCenter as a tier 0 asset, the same tier as a domain controller, and enforce its network boundary outside the appliance where the appliance cannot turn it off. The goal is that only a small, named set of hosts can reach vCenter's management ports at all, and that the appliance cannot reach the open internet.
Concrete steps, in rough order of impact:
- Isolate the management ports. Restrict TCP 443, 22, 5480 and the SSO control port 2012 to a short allowlist of administrative sources, enforced on an upstream firewall or network segment rather than on the appliance.
- Administer from a jump host. Require that management traffic originate from a hardened privileged access workstation or bastion, not from general-purpose servers or user desktops.
- Filter egress. vCenter has no reason to make arbitrary outbound connections. Blocking egress breaks command-and-control and data exfiltration even after a successful intrusion.
- Forward logs off the box. Send appliance and firewall logs to a SIEM so denied connections and shell activity are visible, and so local log deletion does not erase the evidence.
Mandiant published a VCSA hardening and logging approach that enforces these controls at the appliance's Linux layer with iptables, independent of the VAMI, and maps each control to the stage of the attack it would have disrupted. The broader framework lives in Mandiant's vSphere defender guidance. Whatever tool you use, the principle is that network policy for a tier 0 asset must not be modifiable through the asset's own application layer.
Why does this reduce the damage from the next zero-day?
Patching is reactive. It protects you once a vulnerability is known and a fix exists. CVE-2026-123890 was exploited for more than two months before anyone outside the attacker knew it existed, and the exploit tool predated the fix by months. No patch cadence would have helped during that window.
Network isolation works differently. It does not care whether the vulnerability is known. If a compromised internal Linux appliance cannot reach vCenter on port 2012, the exploit payload never arrives, regardless of whether the bug is patched. If forged administrator tokens cannot reach the vSphere API from an untrusted segment, the Golden SAML step fails even though the signing key was stolen. If egress is blocked, the stolen data cannot leave. Defense in depth means each layer assumes the one before it failed.
This is the same reasoning that applies to any management plane, and it connects to why a compromised virtualization management plane hands over every VM it manages. For the specifics of this vulnerability and the patch, see our report on the exploited vCenter authentication bypass CVE-2026-123890, and for the cleanup after exposure, how to audit vCenter SSO accounts after a compromise.