← Lantern Cybersecurity Daily

A vCenter zero-day exploited since September, NGINX2Root patches, and JPMorgan Chase confirms a vendor-side leak (November 20, 2026)

November 20, 2026 · 12m 14s · Listen

If someone gets into the server that runs all your virtual machines, what's left that they can't reach?

The coffee machine, maybe. If it isn't virtualized.

Welcome to Lantern Cybersecurity Daily. It's Friday, November 20. Today: a VMware vCenter zero-day that Mandiant says was exploited back in September, critical NGINX bugs and how exposed you really are, JPMorgan Chase and Mastercard confirming vendor-side leaks, a Russian espionage group using one breach to set up the next, and an exploited flaw in the LiteLLM AI gateway.

Lawrence Abrams, writing in BleepingComputer:

Mandiant observed the tool successfully creating a vCenter Single Sign-On (SSO) administrator account and stealing the server's directory database, which can be used to extract further credentials.

vCenter is the management layer for VMware environments: one server that controls the hypervisors and the virtual machines on them. On Thursday, Broadcom patched CVE-2026-123890, an authentication bypass in vCenter's directory service. An unauthenticated attacker with network access can end up with an administrator session. Broadcom calls it an emergency update and says there's no official workaround, only network restrictions to reduce exposure.

Here's the evidence split. Broadcom hasn't confirmed exploitation. Mandiant and Google's Threat Intelligence Group have. They say a suspected Chinese state-sponsored group they track as Armory Castle, formerly UNC5221, exploited it in early September. That's more than two months before a patch existed.

And it wasn't the first step. Per Mandiant, the group came in through a BeyondTrust Remote Support zero-day, moved to internal Linux appliances that had no endpoint detection, then scanned for VMware ports. Its custom tool, BONESAW, created an SSO admin account and could pull SAML signing keys. Another tool, SAWSNAKE, forges authentication tokens to bypass MFA. They went on to reach ESXi hosts and pull Active Directory database files out of virtual machines.

This one does justify interrupting your day. The product is named, the attacks are confirmed by a credible researcher, and the fix is out. Broadcom says you're affected on any vCenter, SDDC Manager or VMware Tools for Windows below the fixed versions in VMSA-2026-0009, and that includes VMware Cloud Foundation deployments.

The uncomfortable part is the timeline. If you ran an exposed BeyondTrust appliance or vCenter since September, patching closes the door, but it doesn't tell you whether someone already walked through. Google's team published mitigation and hardening guidance alongside its analysis.

BleepingComputer, on how exposed NGINX servers really are:

The two HTTP/2 flaws require an unpatched server with HTTP/2 enabled on a listener an attacker can reach, while CVE-2026-123633 also requires NGINX to proxy requests to a backend.

On Wednesday, F5 shipped patches for 22 vulnerabilities, five of them critical NGINX bugs that SecurityWeek says can lead to unauthenticated remote code execution. The two that earned the name nginx2root are in HTTP/2. CVE-2026-123007, a flow-control memory bug, goes all the way back to version 1.9.5 from 2015. CVE-2026-123633 is a use-after-free when NGINX proxies HTTP/2 traffic to a backend.

Wiz says it got code execution on a server just serving static files over HTTP/2 with TLS. Xint chained the first bug with a Linux kernel flaw in SCTP to escape an NGINX Docker container and get root on the host. Wiz isn't releasing its exploits for now. Censys reproduced the memory corruption in lab tests but didn't reach code execution.

Now the reality check. Xint says 98.9 percent of NGINX hosts that reveal a version are in the affected range. Florian Roth argues that's misleading, because it ignores whether HTTP/2 is on and whether distributions backported fixes. In a small sample he found about two thirds supported HTTP/2. Censys estimates about 8.8 million services potentially exposed to the older bug as of Thursday.

So the headline number is version-string math. Still, two thirds of a very big number is a very big number. The fixed releases are NGINX Open Source 1.30.7 stable and 1.31.9 mainline, and NGINX Plus 37.0.8.1 or 37.3.0.1.

If you can't upgrade right away, the recommendation is to turn off HTTP/2 temporarily, which covers only the two nginx2root bugs. The other three hit HTTP/3 proxying, DNS resolution with subrequests, and mail proxy authentication, so check your config. None of the reports we have describe attacks in the wild yet.

The Associated Press, reporting from New York:

Personal information belonging to thousands of JPMorgan Chase customers was leaked onto the internet Thursday after third-party vendors at the bank were targeted in a cyberattack.

Yesterday we reported that the n0n extortion group had listed JPMorgan Chase, Visa and Mastercard, and that Chase had said a recent outage came from a software bug, not a cyberattack. What's new is that data has now been published and the bank is confirming a vendor-side incident. The AP says the leak included emails and some phone numbers, with no credit card or especially sensitive data, and that Chase calls it unrelated to the outage. TrendWire, citing reporting from The Verge and Cybernews, puts it at about 16,000 customers' names and email addresses.

Mastercard has come around too. It told Cybernews a third-party provider exposed email addresses from a small subset of cardholders in a specific benefit program, with no impact on its payment network. The dataset n0n published for Mastercard is about 2,000 entries, and a Cybernews researcher, Aras Nazarovas, thinks it may be test data. They threatened hundreds of thousands of records and delivered 280 kilobytes.

Visa's leak looks more real. Cybernews says it found more than 6,000 premium Visa customers in the data, with names, phone numbers and dining and travel bookings from Asia Pacific programs. Visa says no card numbers or payment data were affected. And Nazarovas says the files for Chase, Visa and Mastercard share formatting patterns, suggesting one shared third-party source.

That fits the shape Expedia described: one provider's trusted access, many victims. Nobody has named that provider, though, and no one has tied these card-program leaks to Qubole.

So the confirmed part is a vendor exposure of contact and booking data. For anyone who receives these leaked lists, the realistic threat is phishing that cites a real reservation or loyalty program.

Jai Vijayan, writing in Dark Reading, on research from Google's Threat Intelligence Group:

In the second attack, earlier this year, Cold Relic impersonated an employee of the first company to contact an individual at a second organization through an encrypted messaging app.

Cold Relic is Google's name for the group also called Star Blizzard and Callisto, which the US government links to Russia's FSB. In late 2025 it used stolen credentials for a super admin account at a US organization's cloud environment, routed through a residential proxy near the victim, and spammed MFA push prompts until the user approved one.

Then it dug in. SecurityWeek, summarizing Google, says the group registered its own device, added passkeys, rogue MFA devices and backup codes, downloaded the entire cloud drive and exported full employee lists. It held on for months. Even after it was removed, it kept using lookalike accounts to approach the victim's colleagues and contacts.

The second victim got a message from what looked like a known contact, with phishing links styled as documents from their conversation. Cold Relic took that person's Microsoft Entra ID credentials and tokens, came in through federated single sign-on, and bulk-downloaded their Google Workspace files.

What I like here is that the defenses are specific and boring. Google's Joseph Spero recommends phishing-resistant MFA, restricting and auditing passkey registrations, hardening MFA enrollment, auditing OAuth grants, and revoking sessions, not just resetting passwords, when you suspect a compromise.

And a breach doesn't end when the attacker is evicted. Your org chart and email history become someone else's phishing kit, aimed at your partners.

Hazel Stone, in an alert write-up on CISA's latest catalog additions:

CISA added this record based on evidence of active exploitation but did not say what that evidence was.

The record is CVE-2026-40217 in LiteLLM, BerriAI's open-source proxy that sits between applications and model providers. CISA added it to its Known Exploited Vulnerabilities catalog on Wednesday, with a federal deadline of December 2. CISA, the National Vulnerability Database and the GitHub advisory all describe it as code execution via bytecode rewriting at the guardrails test custom code endpoint.

There's a labeling mess, though. The write-up notes that Google's threat intelligence reporting described CVE-2026-40217 differently, as a user ID rebind on a key regeneration endpoint that turns ordinary virtual keys into proxy admin credentials. The author treats that as an unresolved conflict until BerriAI or CISA clarifies.

On scope: the GitHub advisory lists 1.81.8 up to but not including 1.83.10 as affected, and says the default Docker image runs as root. It isn't unauthenticated. In default setups you need a proxy admin credential to reach the endpoint. The recommended fix is 1.83.11 or later.

Needing an admin credential sounds reassuring until you remember we've already covered ARTEX agents hunting AI gateway keys, LiteLLM among them. If those keys leak, so does your protection.

So: upgrade, rotate proxy admin credentials, review who holds them, and keep the management interface off the internet.

If this briefing saves you some scrolling, leave us a rating or a short review wherever you listen. It helps other builders find the show.

Weekend checklist: patch vCenter, check whether HTTP/2 is on in your NGINX configs, and see who holds LiteLLM admin keys.

And if a colleague messages you on a new app, check with them some other way first.

That wraps the week. Lantern Cybersecurity Daily is a Lantern Podcast, and our next episode is Monday, November 23.