< ciso
brief />
NGINX and Java4Shell Flaws Lead a Dense Day of Critical Patches

NGINX and Java4Shell Flaws Lead a Dense Day of Critical Patches

Coverage: 18 Nov 2026 (UTC)

< view all daily briefs >

Two widely deployed software foundations dominated the security agenda on November 18, 2026: the NGINX web server and the OpenJDK Java runtime, each affected by critical flaws that can allow unauthenticated code execution. Active exploitation was confirmed in Perforce Puppet and Apache Roller, and F5 disclosed six critical BIG-IP vulnerabilities that can only be fixed through engineering hotfixes. A hijacked Microsoft PyPI package, two espionage campaigns, a new Amazon Cognito authentication capability and a ransomware arrest operation complete the day.

NGINX and Java4Shell Put Core Web Infrastructure at Risk

A critical flaw in the HTTP/2 support of NGINX, tracked as CVE-2026-123007 and nicknamed nginx2root, has been present in the open-source code since version 1.9.5, released in September 2015, according to The Hacker News. F5 says an unauthenticated attacker can execute code on servers with HTTP/2 enabled where ASLR is disabled or can be bypassed. Wiz researchers bypassed ASLR in tests, and Xint demonstrated code execution with only a reachable HTTP/2 port. Code runs with NGINX worker permissions, but Xint reached root by chaining a separate Linux kernel SCTP bug, even from inside a Docker container; that kernel bug has no CVE or patch yet. Fixes are available in 1.31.9 and 1.30.7, while NGINX Plus R34 to R36 has no fix listed. The same release addresses four further flaws, each scored 9.2 under CVSS 4.0, affecting HTTP/2 backends, resolver subrequests, HTTP/3 and mail proxy deployments. Several of the research teams relied on AI agents or harnesses to find the issues. Shodan shows 56.7 million NGINX hosts, and 98.9% of those disclosing a version fall in the vulnerable range.

No attacks have been reported and the flaw was not in CISA's KEV catalog, but public proof-of-concept code exists. Rapid7 noted that its own attack requires about 128 GB of traffic and specific AVX-512 CPUs. Recommended mitigations include disabling HTTP/2, which breaks services such as gRPC, and blocking the SCTP kernel module. Administrators should perform a full restart after upgrading, since a reload leaves the old executable running. Ubuntu and Debian had not yet shipped fixed packages at the time of reporting.

On the Java side, researchers at Calif disclosed Java4Shell (CVE-2026-114938) on November 17, a memory-corruption flaw in the OpenJDK Base64 decoder scored 9.2 under CVSS 4.0; Oracle rated it 8.1 under CVSS 3.1. As detailed by Cloudflare, the bug sits in an AVX-512 intrinsic, where malformed padding produces a negative length and a write past the output array, allowing an unauthenticated attacker to corrupt heap memory, leak data and achieve remote code execution. Calif demonstrated attacks against Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, and built its first local exploit in about three hours with AI agents. Only systems running OpenJDK 17.0.6 or newer on x86-64 CPUs with AVX-512 VBMI that decode attacker-controlled Base64 are affected; ARM64 is not.

OpenJDK and Oracle shipped patches the same day, and teams must update bundled runtimes and container images and restart services. Because exposure depends on CPU features, the same application can be vulnerable on newer hardware and safe on older servers, which limits the usefulness of conventional software inventories. As temporary mitigations, the startup flags -XX:-UseBASE64Intrinsics or -XX:UseAVX=2 remove the vulnerable code path. Cloudflare added six managed WAF rules covering headers, query values, bodies, path segments, JWT components and WebAuthn values, enabled in Log mode by default, and recommends switching to Block after reviewing matches. A WAF cannot see paths such as outbound TLS responses, and Cloudflare reports no sign of exploitation on its infrastructure after disabling the accelerated routine on its own Java workloads.

Exploited and Critical Flaws in Enterprise Platforms

CISA has added a critical flaw in Perforce Puppet, CVE-2026-115731 (CVSS 9.4), to its Known Exploited Vulnerabilities catalog, citing evidence of active exploitation, according to Puppet KEV report. The deserialization weakness in report-processing code allows a client with any CA-signed agent certificate to submit a crafted report through a PUT request and execute arbitrary code on the Puppet Server as the service user. Where the certificate authority signs requests automatically without strong verification, even an unauthenticated network attacker could obtain a valid certificate. Perforce shipped patches for affected Puppet Enterprise and Puppet Core versions on November 3, 2026, and the open-source OpenVox has also been fixed. Alain Rödel of Neodyme AG reported the bug alongside CVE-2026-115732, a certificate-name authorization bypass rated 9.0, and said the findings were made with AI assistance. Neodyme warned that code execution on a Puppet server could enable theft of CA signing keys, host impersonation, ransomware deployment and lateral movement. Federal Civilian Executive Branch agencies must apply fixes by December 1, 2026, under BOD 26-04.

The Apache Software Foundation warned that CVE-2026-129448 (CVSS 9.8) in Apache Roller is being exploited in the wild, as reported in this Roller advisory coverage. The XML-RPC endpoint in versions 4.0.0 through 6.1.4 accepts vendor extension types and deserializes them before authenticating the requester. ASF stated the issue is the same as CVE-2026-82384, patched in September 2026 but listed at the time as affecting only Roller 6.1.5. Users should upgrade to Roller 6.1.6 or later; those who cannot should set enabledForExtensions to false in WEB-INF/web.xml or remove the /roller-services/xmlrpc servlet mapping, then restart. Disabling XML-RPC in the administration UI is not sufficient, because older versions deserialize the request before checking that setting.

F5 disclosed six critical BIG-IP vulnerabilities that allow an unauthenticated attacker to crash the Traffic Management Microkernel (TMM) and may permit remote code execution, according to BIG-IP flaw report. All six are stack-based buffer overflows in the data plane, rated 9.8 under CVSS v3.1 and 9.3 under CVSS v4.0; the management side is not exposed. Each requires a particular virtual server configuration: five involve iRules using commands such as PEM::subscriber or DNS::scrape, and one involves an APM per-request policy combining the Client Information and Endpoint Security agents. Some also affect BIG-IP Next for Kubernetes and BIG-IP CNF. WatchTowr researcher Sina Kheirkhah reported four of the flaws. The only fixes are engineering hotfixes, which F5 says do not receive the extensive QA of scheduled releases and carry no warranty, and no workaround exists. F5 evaluates only supported versions, so older releases are not proven safe. No exploitation is reported, though an earlier APM flaw, CVE-2026-94127, was exploited and added to the KEV catalog. Six lower-rated BIG-IP and BIG-IQ issues were also addressed, including an 8.8-rated iControl SOAP flaw.

Supply Chain Compromise and State-Aligned Espionage

Attackers briefly hijacked Microsoft's official Agent Framework package on PyPI and used it to distribute credential-stealing malware, as reported by CSO Online. Malicious versions 1.26.1 and 1.26.2 appeared on November 17 and were removed within about an hour. The payload ran every time Python started, even if the package was never imported, and the releases also installed the genuine framework code, so applications kept working. The stealer collected environment variables, up to 60 credential files and secrets from up to 300 running processes, and requested Azure DevOps and managed identity tokens for services such as Key Vault, Microsoft Graph and AI Foundry. Data was sent to a webhook.site address from a module disguised as diagnostics named _telemetry.py. Researchers believe the wheel build process was hijacked, but Microsoft has not said how access was obtained. With roughly 192,000 monthly downloads, any unpinned install during the window, including container images, CI caches and lockfiles, may still contain the malware. Teams should search for the known malicious files, treat every reachable secret as compromised, rotate Azure, GitHub, npm, PyPI, Docker, Hugging Face and OpenAI credentials, pin verified releases with hash checking and block outbound traffic to webhook.site from build infrastructure.

Cisco Talos reported a campaign by UAT-12643, a cluster it assesses with low confidence to be North Korea-aligned, in Talos research. Lures include a financial statement decoy PDF, a trojanized Viber installer and microsip softphone variants, and keyword-based file harvesting in English, French, Spanish and Portuguese suggests a focus on Europe. Initial access relies on weaponized LNK files abusing MpCmdRun.exe and certutil, an HTA run by mshta.exe, and an MSBuild project that compiles C# in memory. The payload, a Go-based Overlord RAT with 35 commands, marks itself as a critical process so forced termination triggers a BSoD, and persists through a Windows Recovery Environment OEM script that survives a PC reset. For C2 resilience it reads encrypted WebSocket URLs from Solana Memo Program transactions, a technique previously linked to Lazarus, while stolen data is uploaded to gofile.io with Discord webhook notifications.

Bitdefender reports that the suspected Chinese APT behind EggStreme has substantially rebuilt its toolkit while continuing to target Philippine government organizations, according to EggStreme coverage. The operators changed the agent's network protocol, expanded its command set with Active Directory reconnaissance, WMI querying and TCP forwarding, and added a passive variant that listens for the operator instead of contacting a C2 server. The keylogger has been removed, and EggStremeLoader moved from RC4 to RC5 and then to a custom cipher by late October 2025. Lateral movement now often uses an unmanaged system to stage payloads launched via DLL sideloading through a VBScript run over SMB. The same report notes Mustang Panda targeting Philippine officials with ASEAN Summit lures to deliver PlugX, and Unit 42's account of Earth Alux, which has attacked Southeast Asian governments since at least July 2025 using HTTP/3-based RATs after entry through unpatched Exchange servers and phishing.

Identity Assurance and Ransomware Arrests

AWS has detailed step-up authentication in Amazon Cognito, described in a AWS Security Blog post. Cognito records the assurance level reached and the methods completed and includes them in issued JWTs through the acr and amr claims defined in OpenID Connect Core. Applications request a target level with acr_values and max_age on managed login, or with equivalent parameters in the USER_AUTH flow, and the feature requires the Essentials or Plus plan. Four default levels range from password only to password plus a TOTP code, and the auth_time claim separates freshness from strength because token refreshes do not reset it. Enforcement remains the responsibility of the application and API, and mapping to frameworks such as NIST SP 800-63 is left to the application.

Ukraine's National Police, working with the FBI, arrested four suspected key members of an international ransomware organization after a two-year investigation, as reported by BleepingComputer. The group of at least 27 people was connected to a wider network of more than 60 members across eight jurisdictions, including former Conti members. At least 28 companies paid over $40.5 million in ransoms, across sectors including oil and gas, pharmaceuticals, manufacturing, logistics, finance and critical infrastructure. The attackers impersonated IT support using a custom platform with AI-generated voice messages, persuading employees to hand over credentials or install TeamViewer and AnyDesk. Court records identify Chaos as the ransomware used; in one case, 1.4 TB was stolen from Warren Oil, which paid roughly $5 million after a $10 million demand. Police conducted 66 searches in Kyiv and 11 regions, seizing devices, nearly $370,000 in cash and €20,000, and the suspects face up to 15 years in prison.