Decade-Old NGINX HTTP/2 Flaw Enables Unauthenticated Code Execution


Nginx

Nginx

Breaking TORONTO, ON

A decade-old vulnerability inside NGINX's HTTP/2 implementation has become one of the biggest web infrastructure emergencies of the year. Disclosed on November 18, 2026 alongside out-of-band patches from F5, the flaw, tracked as CVE-2026-123007, dubbed "nginx2root," and rated 9.2 on CVSS 4.0, allows unauthenticated attackers to execute code on servers with HTTP/2 enabled. The bug has existed in NGINX open source since version 1.9.5 was released on September 22, 2015, and fixed releases 1.31.9 and 1.30.7 shipped on November 18.

Neither F5 nor researchers at Wiz and Rapid7 have reported active exploitation in the wild yet, and CISA's Known Exploited Vulnerabilities catalog did not list it as of November 18. Even so, proof-of-concept exploit code is already on GitHub—though its notes state it is only suited for demonstrating the attack against a freshly started Docker container on the same machine—and researchers have shown working remote code execution (RCE) in tests. F5 says code execution is possible where Address Space Layout Randomization (ASLR) is off or bypassed; Wiz bypassed ASLR and ran code on official NGINX 1.31.7 and 1.31.8 builds over HTTP/2 with TLS on 64-bit systems, while Xint achieved RCE over a reachable HTTP/2 port against test servers on AWS, GCP, and Azure, including over a public IP address.

How CVE-2026-123007 Corrupts Memory

At the code level, CVE-2026-123007 is an integer underflow (CWE-191) inside ngx_http_v2_module, which tracks flow-control credit for outbound HTTP/2 data. Clients grant window credits so the server does not flood the receiver, and NGINX tracks those credits across individual streams as well as the shared connection.

According to researcher breakdowns, an attacker can corrupt memory over an HTTP/2 session through a sequence of stream and credit manipulations:

  1. Queuing and Stream Cancellation: A remote, unauthenticated client slows its reading of server output so outbound DATA frames stay queued while connection credits climb toward maximum bounds and stream credits sit just below that ceiling. Sending a SETTINGS frame that raises SETTINGS_INITIAL_WINDOW_SIZE past the stream limit closes streams while keeping the connection alive, and cleaning up those closed streams can restore reserved credit past maximum limits.
  2. Connection Credit Underflow: Follow-up WINDOW_UPDATE frames and subtraction checks on connection-level counters underflow (CWE-191), pushing the unsigned 64-bit counter into a massive range.
  3. Signed vs. Unsigned Length Mismatch: When NGINX's output filter converts that oversized unsigned counter into a signed 64-bit length (off_t) or buffer calculation, it can produce a negative response length and an invalid buffer state.
  4. Out-of-Bounds memcpy and Code Execution: Buffered TLS output routines can pass a huge unsigned byte count into memcpy, triggering out-of-bounds memory reads and writes. Rapid7 noted its exploit against a 64-bit Linux container running NGINX 1.31.8 with HTTP/2 and TLS required about 128 GB of HTTP/2 traffic over one connection—taking two and a half minutes in its demo, whereas links slower than roughly 300 Mbit/s would run past NGINX's default one-hour connection limit and fail—and relied on a glibc memory-copy routine selected on x86-64 CPUs with avx512f and avx512vl features. Researchers have shown that leaking memory addresses to resolve functions past ASLR and hijacking a cleanup callback allows code execution in the NGINX worker process.

Five Critical Flaws and the Path to Host Root

Code run via CVE-2026-123007 lands with NGINX worker privileges, not host root. According to the Computer Incident Response Center Luxembourg (CIRCL), reaching host root requires a second privilege-escalation step. In Xint's tests, code running as a worker reached host root, even inside a Docker container, by chaining with a separate bug in the Linux kernel's Stream Control Transmission Protocol (SCTP) code.

F5's November 18 release addresses ten vulnerabilities in total, including four other critical flaws rated 9.2 on CVSS 4.0 that can lead to code execution under the same ASLR condition when specific setups are present:

  • CVE-2026-123007: Requires HTTP/2 turned on (via a listen directive with http2 or http2 on;).
  • CVE-2026-123633: Requires HTTP/2 enabled with requests passed to a backend server while a worker is running low on worker_connections.
  • CVE-2026-123634: Requires a proxied server name built from variables plus a Server-Side Include (SSI) or mirror subrequest to a location looking up that same name.
  • CVE-2026-121810: Requires HTTP/3 turned on.
  • CVE-2026-113425: Requires NGINX used as a mail proxy.

What This Means for Canadian Infrastructure

Across Canada, NGINX sits in front of public websites, reverse proxies, API gateways, and cloud workloads across banking, telecom, healthcare, retail, and government. Even when an upstream app server is locked down, compromising an NGINX worker exposes in-memory TLS traffic, session tokens, secrets, and internal network paths.

The Canadian Centre for Cyber Security (CCCS) published advisory AV26-1174 on November 18, 2026, flagging F5's NGINX advisories for Canadian defenders. Critical infrastructure and regulated operators should treat patching and hardening as a priority in line with Bill C-26 expectations and Communications Security Establishment (CSE) / CCCS guidance, while teams subject to PIPEDA or Quebec's Law 25 should factor edge-proxy compromise into incident response and breach assessments with the Office of the Privacy Commissioner of Canada (OPC) or provincial regulators where applicable.

What Defenders Should Do Now

  • Check Exposure and Package Status: Test internet-facing NGINX servers, and test NGINX origins directly behind CDNs, since an origin accepting only HTTP/1.1 cannot be hit over HTTP/2, using openssl s_client -connect <host>:443 -alpn h2 (looking for ALPN protocol: h2). Official Docker images had 1.31.9 and 1.30.7 tags (including Alpine variants) on November 18, whereas Ubuntu (16.04 through 26.10) and Debian (bookworm, trixie, forky, and sid) had not shipped fixed packages as of that date; because distros can backport fixes without changing nginx -v output, compare full package versions against your distribution's advisory.
  • Upgrade and Restart Fully: Per F5, NGINX Open Source 1.31.0 through 1.31.8 (mainline) is fixed in 1.31.9, and 1.9.5 through 1.30.6 (including older branches like 1.28) is fixed in 1.30.7, with both fixing all five flaws. On NGINX Plus, 37.1.0 through 37.2.0.1 is fixed in 37.3.0.1 and 37.0.0 through 37.0.7.1 is fixed in 37.0.8.1 (with R34 through R36 having no fix listed; 37.3.0.1 and 37.0.8.1 fix four of the five flaws, while CVE-2026-123633 was fixed in earlier builds 37.2.0.1 and 37.0.7.1). Perform a full restart or live binary upgrade after updating—a reload alone leaves the old binary running and existing connections on old workers.
  • Apply Workarounds and Block SCTP: CIRCL notes disabling HTTP/2 temporarily blocks CVE-2026-123007 and CVE-2026-123633 (on 1.25.1+, set http2 off; in affected server blocks and remove http2 from listen directives; on 1.9.5 through 1.25.0, remove http2 from listen directives such as listen 443 ssl http2; to listen 443 ssl;), then restart where needed and re-check with openssl s_client. Disabling HTTP/3 addresses CVE-2026-121810, while F5 lists workarounds for CVE-2026-123634 (avoid SSI and mirror directives and do not build proxied server names from variables) and CVE-2026-113425 (restrict mail proxy listeners to trusted networks). For the Linux SCTP flaw, CIRCL advises blocking module loading on hosts not using SCTP via install sctp /bin/false in /etc/modprobe.d/ (which does not unload an already-loaded module or disable built-in kernel SCTP).
  • Watch for Signs of Attack: CIRCL recommends checking for NGINX workers launching unexpected programs, workers crashing abnormally or dropping core dumps, long-lived HTTP/2 connections with heavy inbound traffic and few logged requests, and unusual outbound connections or DNS lookups—noting that clean access logs do not rule out an attack since HTTP/2 control frames typically do not appear as requests there, and no single indicator is proof on its own.

The vulnerabilities in this story

  1. CVE-2026-123007 (cve.org, opens in a new tab)

    Not yet rated by the NVD: see the record on cve.org.

  2. CVE-2026-123633 (cve.org, opens in a new tab)

    Not yet rated by the NVD: see the record on cve.org.

  3. CVE-2026-123634 (cve.org, opens in a new tab)

    Not yet rated by the NVD: see the record on cve.org.

  4. CVE-2026-121810 (cve.org, opens in a new tab)

    Not yet rated by the NVD: see the record on cve.org.

  5. CVE-2026-113425 (cve.org, opens in a new tab)

    Not yet rated by the NVD: see the record on cve.org.

Scores and dates: NVD (not endorsed by the NVD). Each ID opens its record on cve.org.

About the authors

Editor's Picks

The best of cybersecuritynews.ca, delivered to you weekly.

On Display

Critical "Java4Shell" Flaw Threatens Enterprise Infrastructure and Identity Providers

Securing PHP-FPM, session handlers, and memory-safe runtimes on Debian 13 against critical remote code execution flaws.

Deep Dive: PHP’s Seven Critical RCE Flaws, Exploitation Mechanics, and Patching on Debian 13 (Including SURY)

KPMG Canada is notifying affected clients after confirming unauthorised access to client tax documents in a cybersecurity breach.

KPMG Canada Confirms Client Tax Documents Accessed in Cyber Breach