The most useful fact from this week is a sequence. In three of the items below, the attacker was inside before the defender knew there was anything to fix: a vCenter zero-day in use for roughly two months before Broadcom shipped the fix, a vendor's AWS roles used to walk into customer accounts through trust nobody was using, and a LiteLLM flaw fixed in May that only reached CISA's exploited list now. The work this week is to collect evidence, patch, and then hunt, in that order.
Read for a US bank, credit union or fintech security team: what happened, what was actually exploited, and what I would do before next Friday. Every fact links to its source, and anything still developing is marked.
The week at a glance
Status as of November 21:
- Nov 17: CISA urges immediate patching (opens in new tab) of Java4Shell (CVE-2026-114938), fixed in the OpenJDK advisory (opens in new tab). CISA adds (opens in new tab) Puppet Enterprise CVE-2026-115731 to the Known Exploited Vulnerabilities (KEV) catalog.
- Nov 18: CISA adds (opens in new tab) LiteLLM CVE-2026-40217, Synology DSM CVE-2024-10441, REDCap CVE-2026-90817 and Apache Axis CVE-2026-90436 to KEV. NYDFS sends regulated entities an industry letter (opens in new tab) on Java4Shell.
- Nov 19: Broadcom publishes VMSA-2026-0009 (opens in new tab) for vCenter CVE-2026-123890, and Google publishes its analysis (opens in new tab) of the exploitation.
- Nov 20: CISA adds (opens in new tab) vCenter CVE-2026-123890, Cisco Expressway CVE-2026-76466, SAP NetWeaver CVE-2025-42884 and Apache Roller CVE-2026-129448 to KEV. JPMorgan Chase, Visa and Mastercard confirm incidents at third-party providers, as reported by BleepingComputer (opens in new tab) and The Register (opens in new tab).
vCenter: exploited since September, patched this week
vCenter is the console that runs a VMware estate, and this flaw lets anyone who can reach it on the network bypass authentication. Google shows attackers using it to create an SSO administrator account. Broadcom's advisory (opens in new tab) scores CVE-2026-123890 at 9.8, lists fixed builds for the 9.1, 9.0 and 8.0 (U3m) lines and VCF 5.2.5, offers no workaround, and sends 7.0 customers with extended support to Broadcom Support. It does not say whether Broadcom knows of exploitation.
Google's threat intelligence group fills that gap. Its report (opens in new tab) attributes the activity to a suspected espionage actor it tracks as ARMORY CASTLE, and says forensic evidence shows the group's tool, BONESAW, ran against vCenter in early September. The tool copies the vCenter identity database, extracts the key that signs single sign-on tokens, and creates its own administrator account. Google also describes the group pulling the Active Directory database (NTDS.dit) and registry hives out of a guest virtual machine's disk from the hypervisor side. In a bank, that is the step that turns one appliance into the whole Windows domain. Google says the initial access in its earlier report on the same group was a BeyondTrust Remote Support zero-day, CVE-2026-102764, which is also on the KEV catalog (opens in new tab).
A patch applied this week closes the door and says nothing about the two months before it. The patch is not the end of the work. It is the start of the hunt. The indicators Google publishes are specific enough to check: an unexpected local account in the vsphere.local domain, management API calls that switch on SSH for the appliance, and connections to the network indicators Google lists. Google also recommends an egress deny policy on the appliance. It also points to Mandiant's hardening tool for default-deny inbound rules on the management ports, with a warning to test them under formal change control first.
Three vendor incidents, one unconfirmed name
By the end of the week, JPMorgan Chase, Visa and Mastercard had each confirmed an incident at a third-party provider. Chase said in a post (opens in new tab) that "third parties who provide services to the firm were the target of a cyber attack" and that attackers reached "certain marketing data held by those vendors," which its investigation so far puts at names and email addresses. BleepingComputer reports no effect on the bank's services. Visa told BleepingComputer (opens in new tab) that its systems "were not involved and were not impacted" and that "No payments data such as card numbers were affected." Mastercard told the same outlet that a provider's incident exposed a "limited amount of email addresses" from a subset of cardholders in one benefit program. None of the three named the provider.
BleepingComputer reports Qubole as the suspected provider, and SOCRadar told The Register the timing fits the Qubole campaign. Neither link is confirmed, and Wiz says it has not tied its Qubole findings to other reported activity. A group calling itself n0n claimed an attack on Qubole and, per BleepingComputer, posted data it says came from the three firms. Qubole's parent company did not respond to The Register. Researchers at SOCRadar told The Register (opens in new tab) the leaked files hold mobile numbers, card BIN prefixes and dated travel and restaurant bookings, and no card numbers or passwords. SOCRadar's CISO called it what a fraudster needs to pose as a concierge desk or a bank.
Qubole's cross-account roles
The Qubole campaign deserves attention on its own. Wiz reported (opens in new tab) earlier in the month, working with Sentra and Mandiant, that an attacker used credentials tied to Qubole's own AWS account to assume the roles customers had created so Qubole could reach their data. Every request carried the correct External ID, the shared value meant to stop a vendor from being tricked into acting for the wrong customer, and every request succeeded. Many of the assumed roles were not otherwise in use: in most affected environments, the attacker's requests were the only AssumeRole activity from the vendor account in October. The trust policies were simply still there. Once inside, the attacker read Secrets Manager, Parameter Store and Lambda environment variables, then tested what it found against AWS, Entra ID and Okta.
Wiz's remediation note is the line I would put in front of every cloud team: "In this activity, a source IP condition alone would have blocked every role assumption we observed." The second line worth repeating is about cleanup: "Rotating a credential does not end temporary sessions that were already issued." A vendor role is standing access protected by someone else's security program. In vendor management terms, an offboarded vendor with a live trust policy is an offboarding control that never ran.
LiteLLM: the AI gateway holds every key
LiteLLM is an open-source proxy that sits between applications and model providers, so it usually holds the provider API keys for every model a firm uses. Per the GitHub advisory (opens in new tab) for CVE-2026-40217, the /guardrails/test_custom_code endpoint runs user-supplied Python inside a homemade sandbox that "can be escaped using bytecode-level techniques," the proxy runs as root in the default Docker image, and reaching the endpoint requires a proxy-admin credential. That advisory was published in May. Exploitation serious enough for a KEV listing six months later means unpatched proxies are still reachable.
This is the fourth LiteLLM entry in the KEV catalog this year, after additions in May, June and September. A model gateway holds the most valuable credentials in the AI stack, and it belongs in the same patch tier as the VPN concentrator. The GitHub advisory lists 1.83.10 as the patched version while its text names 1.83.11, so I would move to the current release, block the test endpoint at the reverse proxy as the advisory suggests, list every holder of a proxy-admin key, and rotate provider keys on any proxy that was reachable from outside.
Java4Shell: not on KEV, already in a regulator letter
CVE-2026-114938 is a memory-corruption bug in the Base64 decoder of the Java virtual machine on x86-64 processors. CISA's alert (opens in new tab) says crafted Base64 input could crash the JVM and allow unauthenticated remote code execution, and it "strongly urges organizations to immediately patch." The OpenJDK advisory (opens in new tab) scores it 8.1 and marks the 17, 21, 25 and 27 release lines as affected. VulnCheck, as quoted by Help Net Security (opens in new tab), says the vulnerable code runs only on processors with the AVX-512 VBMI instruction set, and turning the crash into code execution takes real exploit work. As of this writing it is not on the KEV catalog, and Help Net Security (opens in new tab) reports no evidence of successful attacks.
It still belongs on a regulated team's list, because NYDFS sent DFS-regulated entities an industry letter (opens in new tab) about it. The letter asks firms to check software inventories for affected Java runtimes, including ones bundled inside third-party applications, container images and appliances, and to work with their software providers to confirm what is installed and what has been mitigated. That is the hard part, because few inventories list the Java runtime inside a vendor appliance. The useful output this week is a list: every runtime found, its version, whether the host processor has the affected instruction set, and who owns the fix. My earlier piece on the PHP fixes argued for setting the patch clock by exposure rather than by score, and Java4Shell is a clean test of that rule.
The three-day clock starts with evidence
Seven of the nine KEV entries added this week carry a federal due date three days out, and six of them carry the catalog's forensic triage flag. Under CISA's BOD 26-04 (opens in new tab), which binds federal executive branch agencies, an entry marked for forensic triage must be remediated within three days and the asset examined to determine whether it was compromised. The implementation guidance (opens in new tab) recommends the order: scope within the first two hours, collect evidence, patch, contain, analyze, and reach an escalation decision by roughly the third day. Its plainest line: "Do not alter or remediate systems prior to evidence/artifact collection when possible."
A bank or credit union is not bound by that directive. The sequence is still worth borrowing for the edge and management appliances that keep showing up on this list. A patch runbook that ends at "patched" leaves the real question open. For any entry flagged for forensic triage, the runbook should start by saving logs and the appliance configuration, and end with a written finding: compromised, not compromised, or not yet known. That finding is what an examiner, an insurer or an incident response firm will ask for, and it fits the decision rights laid out in my first-24-hours incident runbook.
What else hit the catalog
- Perforce Puppet Enterprise, CVE-2026-115731: any client holding a signed agent certificate can run code on the Puppet server, and per the KEV entry, a certificate authority that signs requests automatically lets an unauthenticated attacker get one. A configuration management server is the keys to the fleet.
- Cisco Expressway, CVE-2026-76466: hard-coded credentials on clustered devices that give root command execution. Cisco's advisory (opens in new tab) dates to October and said at publication that Cisco was not aware of malicious use. CISA now lists it as exploited.
- SAP NetWeaver Enterprise Portal, CVE-2025-42884: unauthenticated disclosure or modification of server information.
- Apache Roller, Apache Axis, Synology DSM and REDCap: code execution in all four. For Axis, CISA notes the affected Axis versions could be end of life, which makes a vendor product that embeds Axis the more likely place to find it.
What I would do this week, in order
- vCenter first. Save the appliance logs, then patch to the fixed build for your line. Hunt back to the start of September for the indicators Google lists, and restrict the management ports to an administrative network. Any 7.0 instance without an extended support contract goes to leadership as a risk decision this week.
- Vendor trust policies. List every IAM role that trusts an outside AWS account. Delete the ones for vendors no longer under contract, and add a source IP or VPC endpoint condition to the rest. If a Qubole role exists anywhere, review CloudTrail from October 1 as Wiz recommends, rotate everything that role could read, and revoke its active sessions.
- Brief the fraud and contact center teams. Leaked booking details and card prefixes make impersonation calls convincing. This is where fraud and security share one threat model.
- LiteLLM and any other model gateway. Confirm the version, block the test endpoint, list proxy-admin key holders, and rotate provider keys on anything that was exposed.
- Java runtimes. Inventory them, including inside vendor appliances and container base images. Patch internet-facing services first, and ask each software vendor in writing for its fixed version and date.
- Puppet. Check whether certificate autosigning is on, then patch.
- Write the triage finding. For each forensic-triage entry in your estate, record what was checked, what was found, and who signed off.