Broadcom Patches VMware vCenter Zero-Day as Google Links Months of Attacks to Chinese Espionage Group
Dive Deeper with Our Podcast!
Listen to the Episode: Broadcom Patches VMware vCenter Zero-Day as Google Links Months of Attacks to Chinese Espionage Group
20 November 2026 — Broadcom has issued an emergency fix for a critical flaw in VMware vCenter Server that allows an attacker with network access to take over the management console without a password, hours after Google’s Mandiant unit disclosed that a suspected Chinese state-sponsored group has been exploiting the bug in live intrusions since early September.
The vulnerability, catalogued as CVE-2026-123890, sits in the VMware Directory Service, known internally as vmdird, the identity component at the centre of vCenter. Broadcom’s advisory, VMSA-2026-0009, published late on 19 November, says an unauthenticated attacker who can reach a vulnerable server can bypass login checks entirely and obtain administrative control. Because vCenter manages every virtual machine in a VMware environment, that control extends to the servers those machines run.
Broadcom has classified the update as an emergency release and says no workaround exists, a combination the company reserves for flaws it expects to be attacked at scale. Firewall restrictions can narrow exposure, the company said, but do not remove it. The flaw is present whether or not the Enhanced Linked Mode feature is enabled.
vCenter is the control plane for VMware’s vSphere platform, which by most industry estimates still runs the majority of virtualised workloads inside corporate data centres and private clouds. A single vCenter instance commonly oversees dozens to thousands of virtual machines, including the domain controllers, file servers and line-of-business databases that a company depends on daily. A flaw that hands over the console therefore hands over the business.
Eleven weeks of undetected exploitation
Broadcom’s advisory stops short of confirming that the flaw has been used in attacks. Mandiant and the Google Threat Intelligence Group do not. In a report released the same day, the researchers attribute exploitation beginning in early September to a group they track as Armory Castle, formerly designated UNC5221, which they assess to be operating on behalf of the Chinese state.
The group is already well known to defenders. Its previous campaigns have targeted zero-day flaws in Citrix NetScaler and Ivanti Connect Secure appliances, and most recently a BeyondTrust Remote Support vulnerability, CVE-2026-102764, which has been under attack since late August.
According to Mandiant, the vCenter attacks were not a separate operation but the second stage of that BeyondTrust campaign. After breaching an internet-facing BeyondTrust appliance, the attackers moved onto internal Linux systems, installed web shells the researchers call SLIMSAVAGE, escalated to root and began scanning the internal network for VMware infrastructure.
A custom toolkit built for VMware
Against vCenter, the group deployed a purpose-written exploitation tool in the Go language that Mandiant has named BONESAW. The tool triggers CVE-2026-123890 to open a session with the privileges of vCenter’s built-in top-level administrator, creates a fresh single sign-on administrator account for persistence, and copies out the directory database so that stored credentials can be extracted offline.
BONESAW also retrieves the SAML signing keys that vCenter uses to vouch for user identities to other systems. A second tool, SAWSNAKE, uses those keys to forge authentication tokens, a technique known as Golden SAML. Tokens minted this way arrive already signed, so multifactor authentication is never consulted and any user can be impersonated.
Mandiant’s report lists three further implants observed during the intrusions, CAGEMATCH, BEARBONES and LAZYBONES, used to run commands, transfer files, harvest credentials and maintain access. The attackers ultimately reached the ESXi hypervisor hosts beneath vCenter and pulled data straight out of the virtual machines. Among the files taken were the Active Directory databases from Windows domain controllers, which hold the password hashes for every account in an organisation.
Two zero-day flaws used in one campaign, with malware tailored to each platform encountered along the way, points to a well-resourced actor, the researchers said.
The sequencing is what distinguishes this operation from opportunistic scanning. The group did not find vCenter exposed on the internet and attack it directly. It entered through a remote-support appliance, established a foothold on ordinary Linux servers, mapped the internal network, and only then turned its VMware-specific tooling on the management console. Each stage used a different implant written for the platform in front of it. Mandiant describes the result as access to the victims’ virtualised environments at a depth that allowed data to be lifted from inside the virtual machines themselves, rather than from the network around them.
Who is affected
Broadcom’s scope statement is broad. Every release of vCenter Server, of SDDC Manager, and of the Windows guest agent older than the fixed builds listed in the advisory is affected, whether deployed on its own or as part of VMware Cloud Foundation, vSphere Foundation, Telco Cloud Platform or Telco Cloud Infrastructure. The inclusion of the Windows guest agent means the update reaches into every Windows virtual machine that carries it, not only the management servers.
The same advisory resolves four additional vulnerabilities across SDDC Manager and the Windows guest agent. None of those has been reported as exploited.
The affected population includes a large share of small and mid-sized businesses that run VMware without knowing it, because a managed service provider or hosting company built their environment. Those businesses are exposed to the same flaw; the difference is that someone else holds the patch button.
A familiar pattern from a familiar actor
Armory Castle’s method has been consistent across its known campaigns: identify an enterprise appliance that sits at the boundary of a corporate network, find a flaw in it before the vendor does, and use the resulting foothold to go deeper. Citrix NetScaler and Ivanti Connect Secure were the previous targets of record; BeyondTrust Remote Support is the newest. vCenter marks a shift inward, from the perimeter devices that face the internet to the infrastructure that sits behind them. Security researchers who track the group have noted that its interest lies in long-term access and data collection rather than disruption, which is consistent with the credential and directory theft Mandiant documented here.
The disclosure also continues a difficult year for VMware customers. Since Broadcom completed its acquisition of VMware, the product line has been through licensing changes, support-model revisions and a steady cadence of security advisories, and a number of organisations have been weighing whether to migrate to other hypervisors or to cloud platforms. An actively exploited authentication bypass in the core management product will sharpen those conversations, though for the moment the practical question for every VMware customer is the same regardless of long-term plans: whether the patch has been applied.
Why the management console is the prize
Security practitioners have long treated vCenter as one of the most consequential systems in a corporate network, and this campaign shows why. The console can start, stop, clone, snapshot and read the disks of every server it manages. An attacker with vCenter control has no need to break into a file server, a database server or a domain controller individually; copying their virtual disks and reading them offline achieves the same result, as the Active Directory theft in this case demonstrates. The same access allows snapshots to be deleted before encryption, which is why ransomware groups have pursued vCenter and ESXi for years.
The identity dimension raises the stakes further. Because vCenter’s directory service issues tokens that other systems trust, theft of its signing keys compromises not only the VMware estate but anything downstream that accepts those tokens.
What defenders are being told
Both Broadcom and Mandiant are urging immediate patching. Mandiant adds that organisations should treat the period from September onward as a window of possible compromise and search for its indicators: single sign-on administrator accounts nobody created, unexpected administrator sessions in the directory service logs, and the named web shells on internal Linux hosts. Where compromise is suspected, the researchers advise rotating vCenter’s SAML signing certificate and administrator credentials and treating everything in the directory database as exposed.
The attack path also renews attention on network placement. In the intrusions Mandiant documented, vCenter was reached from inside the network after the initial breach through a remote-access appliance, which means keeping the console off the internet is necessary but not sufficient. Isolating it on a dedicated management network remains the standard cybersecurity recommendation.
Organisations using BeyondTrust Remote Support are being told to confirm they are patched for CVE-2026-102764, the first link in the chain.
For businesses whose VMware environment is operated by a third party, the guidance translates into questions for the provider: whether every vCenter instance in the environment has been updated, whether anyone has searched for indicators of compromise dating back to September, and whether the console is isolated from the networks that end users sit on. Providers who can answer with specifics and dates are doing the work; those who cannot are the ones to press.
Backup posture is also under scrutiny. An attacker with vCenter control can delete snapshots and backup virtual machines along with the production ones, so copies that live inside the VMware estate offer no protection in this scenario. Immutable backups stored outside vCenter’s reach, restorable without trusting the hypervisor, are the standard defence that practitioners are repeating this week.
Broadcom’s silence on exploitation
One detail in the disclosure has drawn comment from the security community. Broadcom’s advisory does not state that the flaw has been exploited, while Mandiant’s report, published the same day, documents exploitation in detail with named tooling. The discrepancy is not unusual; vendors frequently decline to confirm attacks until their own investigation concludes, and researchers who respond to intrusions often have evidence the vendor does not. But it means that administrators reading only the vendor advisory would see a critical patch without the context that it has already been weaponised for eleven weeks. Practitioners are advising colleagues to read both documents.
Timing
The advisory and the Mandiant report both landed on a Wednesday evening in the United States, leaving many IT teams to read them on Thursday and decide on Friday whether to patch before the weekend. With public disclosure now complete, security researchers routinely expect criminal groups to reverse-engineer a fix within days, and vCenter flaws have historically been adopted quickly by ransomware operators once details are public.
How the industry has responded
Within hours of the two disclosures, managed service providers and hosting companies that operate VMware on behalf of customers began issuing their own notices, several of them committing to patch windows over the weekend of 21 and 22 November. Security vendors that sell detection products for virtualised environments have started publishing rules for the implant names Mandiant released. Cyber-insurance carriers, which in recent policy cycles have added hypervisor controls to their underwriting questionnaires, are expected to treat an unpatched vCenter after this week as a material fact in any subsequent claim.
Independent researchers have also begun examining the patch to understand the underlying bug, a routine step that typically precedes the appearance of public proof-of-concept code. Once that code circulates, the pool of attackers widens from a single state-backed group to anyone with a scanner and a target list, which is the point at which ransomware affiliates have historically moved on VMware flaws. The interval between vendor patch and criminal exploitation for comparable vCenter vulnerabilities in recent years has ranged from days to a few weeks.
What to watch
- Whether Broadcom updates its advisory to confirm exploitation or publish indicators of compromise.
- Whether the US Cybersecurity and Infrastructure Security Agency adds CVE-2026-123890 to its Known Exploited Vulnerabilities catalogue, which would impose a patch deadline on federal agencies.
- The appearance of public exploit code.
- Further reporting from Mandiant or Google on the sectors and regions targeted.
Broadcom did not respond to a request for comment on whether it has observed exploitation in its own telemetry. Mandiant said it would publish additional indicators as its incident-response engagements conclude. This report will be updated as both companies release further details.
Source: BleepingComputer, Lawrence Abrams, 19 November 2026. Primary sources: Broadcom security advisory VMSA-2026-0009 and the Mandiant / Google Threat Intelligence Group report of 19 November 2026.
