On November 18, 2026, NGINX released 1.30.7 (stable) and 1.31.9 (mainline), and F5 published fixes for a set of memory corruption flaws in the NGINX worker process. The most important are in the HTTP/2 code: CVE-2026-123007, an integer underflow in HTTP/2 flow control that leads to a heap buffer overflow. It has been in NGINX since HTTP/2 support arrived in 1.9.5 in 2015. nginx.org rates it major. CVE-2026-123633, a use-after-free in HTTP/2 connection handling, reachable when NGINX proxies HTTP/2 client requests to a backend and the worker is running low on connection slots. It affects 1.19.7 and later. nginx.org rates it medium. The same releases fix further use-after-free flaws that need less common configurations: CVE-2026-123634 (resolver, with SSI or mirror subrequests), CVE-2026-121810 (HTTP/3) and CVE-2026-113425 (mail proxy). F5 scores all of them 8.1 under CVSS 3.1 and 9.2 under CVSS 4.0, with high attack complexity, and says each can lead to code execution “when ASLR is disabled or when the attacker can bypass ASLR.” The research firm Xint calls the set nginx2root. Xint, Wiz Research, Rapid7, Ikotas Labs and other researchers reported the bugs, several of them independently. Rapid7 published a working exploit for CVE-2026-123007 on the day of disclosure. No attacks have been reported, and none of the CVEs is in CISA’s Known Exploited Vulnerabilities catalog. CVE CVSS EPSS CISA KEV Fixed in CVE-2026-123007 8.1 (CVSS 3.1), 9.2 (CVSS 4.0), F5 Not yet scored No NGINX 1.30.7 or 1.31.9, NGINX Plus 37.0.8.1 or 37.3.0.1 CVE-2026-123633 8.1 (CVSS 3.1), 9.2 (CVSS 4.0), F5 Not yet scored No NGINX 1.30.7 or 1.31.9, NGINX Plus 37.0.7.1 or 37.2.0.1 CVE-2026-123634 8.1 (CVSS 3.1), 9.2 (CVSS 4.0), F5 Not yet scored No NGINX 1.30.7 or 1.31.9, NGINX Plus 37.0.8.1 or 37.3.0.1 CVE-2026-121810 8.1 (CVSS 3.1), 9.2 (CVSS 4.0), F5 Not yet scored No NGINX 1.30.7 or 1.31.9, NGINX Plus 37.0.8.1 or 37.3.0.1 CVE-2026-113425 8.1 (CVSS 3.1), 9.2 (CVSS 4.0), F5 Not yet scored No NGINX 1.30.7 or 1.31.9, NGINX Plus 37.0.8.1 or 37.3.0.1 CVSS and EPSS from NVD and FIRST, exploitation status from the CISA Known Exploited Vulnerabilities catalog, checked November 19, 2026. What is confirmed Point Status Fixed in NGINX 1.30.7 and 1.31.9, released November 18, 2026 nginx.org CVE-2026-123007 affects 1.9.5 to 1.31.8, CVE-2026-123633 affects 1.19.7 to 1.31.8 nginx.org security advisories Both need HTTP/2 enabled. CVE-2026-123633 also needs NGINX proxying to a backend F5 Code execution runs as the NGINX worker user, not root F5, Rapid7, Xint Public exploit for CVE-2026-123007 Rapid7, against a fresh NGINX 1.31.8 Docker container Public exploit for CVE-2026-123633 None. Wiz and Xint demonstrated it privately Exploitation in the wild None reported. Not in CISA KEV on November 19 Distribution packages Alpine fixed. Ubuntu, Debian, Red Hat and SUSE pending on November 19 How the flaws work CVE-2026-123007. HTTP/2 limits how much data each side may send with a flow-control window, which the client raises with WINDOW_UPDATE frames. When a stream is cancelled while its response is still queued, NGINX hands the reserved credit back to the connection window without checking the protocol maximum. A later check on WINDOW_UPDATE uses an unsigned calculation that wraps, so the client can keep raising the window. Pushed far enough, the counter turns negative when NGINX converts it for sending, and the TLS send path copies data with a huge length, reading and writing past the end of a heap buffer. F5 says this is “typically followed by an NGINX worker segmentation fault.” CVE-2026-123633. NGINX marks an idle HTTP/2 connection as reusable before checking whether a HEADERS frame is still half received. When the rest of the frame arrives, the connection starts a request but stays on the reusable list. If that request opens a connection to a backend while the worker is short of connection slots, NGINX may close the client connection to free a slot, and the request’s memory is freed while the backend code still uses it. The resolver, HTTP/3 and mail proxy flaws follow the same pattern of memory used after it is freed, in the resolver (CVE-2026-123634), the HTTP/3 module (CVE-2026-121810) and the mail proxy (CVE-2026-113425). Each needs a configuration most web servers do not have. Practical exploit effort The 9.2 score describes what happens when an exploit works. It does not describe how much work the exploit takes, and for these flaws that differs a lot from one CVE to the next. CVE-2026-123007 has a working public exploit, with heavy conditions. Rapid7’s code runs a command on a stock NGINX 1.31.8 Docker image over HTTPS with HTTP/2, and in Rapid7’s demonstration it took two and a half minutes. To get there it needs: About 128 GB of traffic on one connection. The flow-control counter has to be pushed past 2 to the power of 63, which takes billions of WINDOW_UPDATE frames. Rapid7 puts the transfer at about 2 minutes at 10 Gbit/s and about 17 minutes at 1 Gbit/s. NGINX closes a connection after one hour by default (keepalive_time), so an attacker with less than about 300 Mbit/s to the server cannot finish. A specific processor and C library path. Rapid7’s exploit relies on the glibc memcpy variant chosen on x86-64 processors with AVX-512. Nobody has shown it working on ARM64 or on musl-based builds such as Alpine. A quiet worker. Rapid7 states that its public code is only suitable for a freshly started container and will likely fail against a worker that has served traffic for a while, with other clients connected, with too little bandwidth or with a proxy in front. Xint says its own exploit works under normal concurrent traffic against public HTTPS endpoints on AWS, Google Cloud and Azure, but it has not released that exploit. Address space layout randomization (ASLR) is not the barrier F5’s wording suggests. The same bug leaks heap and library addresses in the TLS response, and both Rapid7 and Xint use that leak to get past ASLR. CVE-2026-123633 is harder and has no public exploit. Wiz and Xint reached code execution in their labs and kept the code. An attacker needs: a server that proxies HTTP/2 client requests to a backend with proxy_pass, grpc_pass or a similar directive the worker nearly out of connection slots, about 960 of the default 1,024 worker_connections in use, which the attacker fills by holding connections open precise timing between a half-sent HEADERS frame, a TCP reset and the backend connection, and in Xint’s chain, NGINX opening a connection to a host the attacker controls The resolver, HTTP/3 and mail proxy flaws need a matching configuration first. Xint describes chains for them but withheld the final steps, and no public code exists. Crashing a worker is the easy part. A failed attempt is likely to crash the worker. The NGINX master process starts a new one, and Xint notes the replacement has the same memory layout, so an attacker can retry. Even failed attempts therefore show up as worker crashes and dropped connections. A working exploit does not give root. Code runs as the worker user, usually www-data or nginx. From there an attacker can read the TLS private keys in the worker’s memory, traffic passing through the proxy in clear text, backend credentials in the configuration and anything else the worker account can read. Xint reached root by chaining a separate Linux kernel flaw in the SCTP protocol, present since kernel 4.15 and unpatched when Xint published, and used the same chain to escape a Docker container. Our assessment of the effort, as of November 19: CVE-2026-123007 CVE-2026-123633 Resolver, HTTP/3, mail proxy flaws Public exploit Yes, for a lab target No No Crash a worker Easy, with enough bandwidth Moderate, needs slot exhaustion Needs the matching configuration Code execution on a busy production server Skilled work: the public code must be adapted to the target’s build, processor and traffic Specialist work: a private chain and precise timing Specialist work, configuration permitting Most exposed Internet-facing NGINX on x86-64 reachable directly at high bandwidth Reverse proxies under heavy load Servers with those features turned on Loudest signal About 128 GB on a single connection Hundreds of held connections from one source Worker crashes For a capable exploit developer, adapting Rapid7’s code to a common distribution build is a matter of days, not months. For most attackers today, the realistic outcome is crashed workers and dropped connections. That changes quickly once someone publishes a version that works under real traffic, so patch on the assumption that it will happen. Who is affected, in priority order Internet-facing NGINX with HTTP/2 over TLS, reachable directly rather than through a CDN, on versions 1.9.5 to 1.31.8. HTTP/2 is a common performance setting, so assume it is on until you have checked. NGINX reverse proxies and API gateways that pass HTTP/2 client traffic to backends, which adds CVE-2026-123633. Kubernetes ingress controllers built on NGINX. The community ingress-nginx project was archived in March 2026, ships an affected NGINX build with HTTP/2 on by default, and has no fixed release. F5’s NGINX Ingress Controller 5.8.0, released November 18, bundles the fixed NGINX. Products and appliances that embed NGINX, including hosting control panels, application stacks and container images that carry their own copy. The operating system’s package manager does not update these. Servers behind an L7 CDN or load balancer that terminates HTTP/2, such as Cloudflare, CloudFront or an AWS Application Load Balancer. Rapid7 expects such a proxy to break the CVE-2026-123007 attack, because it does not pass the client’s WINDOW_UPDATE frames through, and Wiz says it may mitigate parts of the exploit depending on the implementation. Nobody has published a test. It protects the origin only if clients cannot reach the origin directly. A layer 4 load balancer, such as an AWS Network Load Balancer, passes the attack through. NGINX Plus customers on 37.0.7.1 (released October 14) or 37.2.0.1 already had the CVE-2026-123633 fix, but still need 37.0.8.1 or 37.3.0.1 for the rest. Affected and fixed versions CVE Feature it needs NGINX Open Source affected NGINX Plus fixed CVE-2026-123007 HTTP/2 1.9.5 to 1.31.8 37.0.8.1, 37.3.0.1 CVE-2026-123633 HTTP/2 and a proxied backend 1.19.7 to 1.31.8 37.0.7.1, 37.2.0.1 CVE-2026-123634 Variable proxy server name with SSI or mirror 0.6.18 to 1.31.8 37.0.8.1, 37.3.0.1 CVE-2026-121810 HTTP/3 1.25.0 to 1.31.8 37.0.8.1, 37.3.0.1 CVE-2026-113425 Mail proxy 0.5.33 to 1.31.8 37.0.8.1, 37.3.0.1 All are fixed in NGINX Open Source 1.30.7 and 1.31.9. NGINX Plus R33 is affected and has no fix listed, so move to R37. The same releases also fix CVE-2026-118899, a TLS 1.3 infinite loop that nginx.org rates major, and several minor flaws. Distribution status on November 19, 2026: Distribution Status Alpine Fixed in nginx 1.30.7-r0 in edge and 3.24 Ubuntu All supported releases affected, from 16.04 to 26.10. No fixed package yet Debian 12 (1.22.1), 13 (1.26.3) and testing (1.30.6) affected. No fixed package yet Red Hat Under review, CVSS 3.1 8.1 SUSE Under analysis, rated important Distributions often backport fixes without changing the version number, so nginx -v may still show an old version after a fixed package is installed. Confirm against your distribution’s advisory. How to check a server These commands only read state. Run them on each server or inside each container that runs NGINX: nginx -v nginx -V 2>&1 | tr ' ' '\n' | grep -E 'http_v2|http_v3|mail' nginx -T 2>/dev/null | grep -nE '^\s*(listen|http2|http3|quic)\b' nginx -T 2>/dev/null | grep -nE '^\s*(proxy_pass|grpc_pass|fastcgi_pass|uwsgi_pass|resolver|ssi|mirror|mail|worker_connections|keepalive_time)\b' HTTP/2 is on when a listen line includes http2 or a server block has http2 on;. The second grep shows whether the server proxies to backends and uses the resolver, SSI, mirror or mail features. Check the package and the kernel module the root chain uses: dpkg -l 'nginx*' 2>/dev/null | grep ^ii rpm -qa 'nginx*' 2>/dev/null apk info -v nginx 2>/dev/null lsmod | grep -w sctp For containers and Kubernetes, list the NGINX images in use: docker ps | grep -i nginx kubectl get pods -A -o jsonpath='{range .items[*]}{.spec.containers[*].image}{"\n"}{end}' | grep -i nginx | sort -u Mitigation and remediation Do these in this order: Upgrade to NGINX 1.30.7 or 1.31.9, NGINX Plus 37.0.8.1 or 37.3.0.1, or your distribution’s fixed package, then restart NGINX so the old binary leaves memory. Where no fix is available yet, turn HTTP/2 off. On NGINX 1.25.1 and later, add http2 off; to each affected server block. On older versions, remove http2 from the listen lines. Test and reload with nginx -t && nginx -s reload. This blocks CVE-2026-123007 and CVE-2026-123633, but not the resolver, HTTP/3 or mail proxy flaws. Servers that front gRPC need HTTP/2 and cannot use this workaround. For the community ingress-nginx controller, set use-http2: "false" in the controller ConfigMap and plan a move to a maintained controller, since no fix will come. Shorten the connection lifetime. The CVE-2026-123007 exploit needs about 128 GB on one connection. With keepalive_time 10m; an attacker would need about 1.7 Gbit/s sustained on a single connection. This reduces the risk and does not fix the flaw. Apply F5’s workarounds for the less common flaws. Avoid SSI and mirror directives where the proxied server name uses variables (CVE-2026-123634). Restrict mail proxy listeners to trusted networks (CVE-2026-113425). Turn HTTP/3 off where you do not need it (CVE-2026-121810). Close the route to root. Unless a host needs SCTP, stop the module from loading with echo "install sctp /bin/false" > /etc/modprobe.d/disable-sctp.conf. This does not help where SCTP is built into the kernel. Limit what NGINX can reach. Restrict outbound connections from NGINX hosts to known backends, and use limit_conn to cap connections per client address, which makes filling worker slots from one source harder. Keep origins behind the CDN. Where an L7 CDN or load balancer sits in front, allow connections to the origin only from its addresses. Free SOC trialWant this watched across your environment without building the hunts yourself? Our 24x7 SOC monitors your endpoints, servers and cloud during a free trial.Start a free trialThreat hunting No attacks have been reported and there are no indicators of compromise. Hunting relies on where vulnerable NGINX runs, the traffic each exploit needs, and what a compromised worker does next. The CVE-2026-123007 exploit is loud. About 128 GB on a single HTTPS connection to a web server has almost no legitimate source, and firewall or load balancer session logs can show it. Failed attempts crash workers, which NGINX records in its error log as worker process <pid> exited on signal 11, and the kernel logs as a segfault. Example KQL for Microsoft Defender and Microsoft Sentinel The KQL below is example code for Microsoft Defender for Endpoint advanced hunting and Microsoft Sentinel. Adapt table and field names to your environment and test each query before relying on it. Hunt 1: where NGINX runs // Example: Defender vulnerability management inventory, NGINX by version DeviceTvmSoftwareInventory | where SoftwareName has "nginx" | summarize Devices=dcount(DeviceId), DeviceList=make_set(DeviceName, 50) by SoftwareVendor, SoftwareName, SoftwareVersion | order by Devices desc The inventory misses copies bundled with products. This finds them by the path of the running binary: // Example: Defender for Endpoint, every nginx binary that has run, with its path DeviceProcessEvents | where Timestamp > ago(14d) | where FileName =~ "nginx" | summarize Starts=count(), LastSeen=max(Timestamp), SampleCommand=any(ProcessCommandLine) by DeviceName, FolderPath, ProcessVersionInfoProductVersion | order by DeviceName asc Hunt 2: single connections carrying huge volumes // Example: Sentinel CommonSecurityLog from a firewall or load balancer, // sessions to HTTPS that moved more than 20 GB CommonSecurityLog | where TimeGenerated > ago(14d) | where DestinationPort == 443 | extend TotalBytes = tolong(SentBytes) + tolong(ReceivedBytes) | where TotalBytes > 20000000000 | summarize Sessions=count(), MaxGB=round(max(TotalBytes) / 1e9, 1), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated) by SourceIP, DestinationIP, DeviceVendor | order by MaxGB desc Backups, replication and large downloads can match. A single client sending tens of gigabytes to a public web server is the case to look at. Hunt 3: NGINX worker crashes // Example: Sentinel Syslog, NGINX workers killed by a signal and kernel segfaults in nginx Syslog | where TimeGenerated > ago(14d) | where (SyslogMessage has "worker process" and SyslogMessage has "exited on signal") or (SyslogMessage has "segfault" and SyslogMessage has "nginx") or SyslogMessage has "worker_connections are not enough" | summarize Events=count(), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), Sample=any(SyslogMessage) by Computer | order by Events desc NGINX writes worker crashes to its own error log unless error_log is sent to syslog, so check that the error log reaches your SIEM. The worker_connections are not enough message can point to the slot exhaustion CVE-2026-123633 needs. A burst of crashes on an internet-facing server is worth treating as a possible exploitation attempt. Hunt 4: NGINX starting a shell or a download tool // Example: Defender for Endpoint, processes started by nginx DeviceProcessEvents | where Timestamp > ago(14d) | where InitiatingProcessFileName =~ "nginx" | where FileName in~ ("sh", "bash", "dash", "zsh", "busybox", "python", "python3", "perl", "curl", "wget", "nc", "ncat", "socat", "base64", "chmod") | project Timestamp, DeviceName, AccountName, InitiatingProcessCommandLine, FileName, ProcessCommandLine | order by Timestamp desc NGINX does not start child processes in normal operation, other than new workers, so any result here deserves a look. Hunt 5: new outbound connections from NGINX // Example: Defender for Endpoint, public destinations nginx contacted this week but not in the prior 30 days let Baseline = DeviceNetworkEvents | where Timestamp between (ago(37d) .. ago(7d)) | where InitiatingProcessFileName =~ "nginx" | where ActionType == "ConnectionSuccess" and RemoteIPType == "Public" | distinct DeviceName, RemoteIP; DeviceNetworkEvents | where Timestamp > ago(7d) | where InitiatingProcessFileName =~ "nginx" | where ActionType == "ConnectionSuccess" and RemoteIPType == "Public" | join kind=leftanti Baseline on DeviceName, RemoteIP | summarize Connections=count(), FirstSeen=min(Timestamp), Ports=make_set(RemotePort, 10) by DeviceName, RemoteIP, RemoteUrl | order by FirstSeen desc A reverse proxy talks to a fixed set of backends. A new public destination may be a changed backend, or a compromised worker calling out. Responding to a suspected compromise Keep the evidence before restarting anything: the NGINX error and access logs, any core dumps, load balancer session logs and, if you can, a memory image of the NGINX processes. Isolate the host from untrusted networks, upgrade NGINX or turn HTTP/2 off, and restart it. Review what the worker account did after the first crash: child processes, new files in web roots and temporary directories, cron entries and outbound connections. Treat the host as root-compromised if the SCTP module was loaded and the kernel is unpatched, and rebuild it. Replace TLS certificates and private keys, since the worker holds them in memory, and rotate backend credentials, API keys and anything else in the NGINX configuration. Timeline September 2015: NGINX 1.9.5 adds HTTP/2 support, the start of the CVE-2026-123007 range. October 11, 2026: Xint finds the flaws, and reports them to F5 on October 16. October 14, 2026: NGINX Plus 37.0.7.1 fixes CVE-2026-123633. October 21, 2026: Wiz Research reports CVE-2026-123633. F5 tells Wiz another team reported it first. November 18, 2026: NGINX 1.30.7 and 1.31.9 are released and F5 publishes the CVEs. Rapid7 publishes a technical write-up and a working exploit for CVE-2026-123007. Xint and Wiz publish their write-ups. F5 releases NGINX Plus 37.0.8.1 and 37.3.0.1 and NGINX Ingress Controller 5.8.0. November 19, 2026: No exploitation reported, and not in CISA KEV. Alpine has fixed packages. Ubuntu, Debian, Red Hat and SUSE are pending. Frequently Asked Questions Is CVE-2026-123007 being exploited? No exploitation had been reported by November 19, 2026, and it is not in the CISA KEV catalog. Rapid7’s exploit is public, so expect attempts against exposed servers. How hard is nginx2root to exploit in practice? Crashing a worker is easy for CVE-2026-123007 given enough bandwidth. Running code on a busy production server takes skilled work: the public exploit needs about 128 GB on one connection, an AVX-512 x86-64 processor and a quiet worker, and has to be adapted to each build. CVE-2026-123633 has no public exploit and needs precise timing. Which NGINX versions are affected? CVE-2026-123007 affects NGINX 1.9.5 to 1.31.8 with HTTP/2 on, and CVE-2026-123633 affects 1.19.7 to 1.31.8 with HTTP/2 on and a proxied backend. Upgrade to 1.30.7 or 1.31.9, or NGINX Plus 37.0.8.1 or 37.3.0.1. Does turning off HTTP/2 protect us? It blocks CVE-2026-123007 and CVE-2026-123633, which matter most for web servers. It does not block the resolver, HTTP/3 or mail proxy flaws, and it is not an option for gRPC frontends. Use http2 off; on NGINX 1.25.1 and later. We are behind Cloudflare, CloudFront or an AWS load balancer. Are we safe? An L7 proxy that terminates HTTP/2 is expected to break the CVE-2026-123007 attack, but nobody has tested it, and it only helps if the origin cannot be reached directly. A layer 4 load balancer does not help. Patch the origin either way. Is ingress-nginx in Kubernetes affected? Yes. The community project is archived, has HTTP/2 on by default and will not get a fix. Set use-http2: "false" in its ConfigMap and plan a move to a maintained controller. F5’s NGINX Ingress Controller 5.8.0 includes the fix. Does the exploit give an attacker root? Not on its own. Code runs as the NGINX worker user. Xint reached root by chaining a separate Linux kernel SCTP flaw, so block the SCTP module where you do not need it. Why does nginx -v still show the old version after we patched? Distributions often backport security fixes into their existing version. Check the package version against your distribution’s advisory instead. How ThreatDefence is responding ThreatDefence has added CVE-2026-123007, CVE-2026-123633 and the related NGINX CVEs to our vulnerability intelligence. Across all monitored environments, our SOC is: identifying internet-facing NGINX servers, reverse proxies and Kubernetes ingress controllers with HTTP/2 enabled watching for very large single HTTPS sessions, NGINX worker crashes, and NGINX processes that start shells or open new outbound connections helping customers confirm the fixed version or the HTTP/2 workaround on each server If you want help finding NGINX across your environment or investigating a suspicious crash, contact us. Read more about our managed detection and response. Sources nginx.org, nginx security advisories (November 18, 2026) nginx.org, Changes with nginx 1.31.9 (November 18, 2026) nginx.org, Changes with nginx 1.30.7 (November 18, 2026) NVD, CVE-2026-123007 (November 18, 2026) NVD, CVE-2026-123633 (November 18, 2026) Rapid7, CVE-2026-123007: critical NGINX HTTP/2 integer underflow leads to RCE (November 18, 2026) Rapid7, CVE-2026-123007 proof of concept on GitHub (November 18, 2026) Xint, nginx2root: remote code execution vulnerabilities in nginx (November 18, 2026) Wiz Research, Unauthenticated remote code execution in NGINX (November 18, 2026) NGINX Plus releases NGINX Ingress Controller releases Kubernetes, ingress-nginx repository (archived) Ubuntu, CVE-2026-123007 Debian security tracker, CVE-2026-123007 Alpine security tracker, CVE-2026-123633 CISA, Known Exploited Vulnerabilities catalog (checked November 19, 2026)