Newshot

Two NGINX HTTP/2 Vulnerabilities Allow Unauthenticated Remote Code Execution

AI summary

Five independent research teams, including Wiz, discovered two memory-safety vulnerabilities in NGINX's HTTP/2 processing path, both rated 9.2 under CVSS 4.0, allowing attackers to execute code on the server without any account. F5 released fixed versions on November 18; until upgrades are complete, temporarily disabling HTTP/2 can close the door. The flaws are a use-after-free (CVE-2026-123633) and an integer underflow (CVE-2026-123007).

Translation(model translation, not the source's own text)

Two NGINX HTTP/2 Vulnerabilities: Unauthenticated Code Execution on the Server

Wiz and other teams discovered two HTTP/2 memory safety vulnerabilities in NGINX, CVSS 9.2, allowing code execution on the server without authentication; one of them dates back as far as version 1.9.5 in 2015. F5 released fixes on November 18; until you upgrade, you can disable HTTP/2 first.

Cover image from Wiz Research (direct link to official site)
Cover image from Wiz Research (direct link to official site)
  • Two Critical vulnerabilities allow attackers to execute arbitrary code without logging in
  • Deployments with HTTP/2 enabled are affected; one vulnerability dates back to 2015
  • The memory safety flaws went hidden for over a decade; AI scanning is what dug them out
  • F5 released patches on November 18; if you can't upgrade, disable HTTP/2 first

The source is the research blog of security vendor Wiz; the vulnerabilities were independently discovered by Wiz together with Ikotas Labs, Metnir, Rapid7, and Xint, and reported to F5. The PoC and impact assessment in this article come from Wiz's own testing; no third-party retest data has been seen.

You open your browser, type a commonly used URL, and the page loads as usual. The machine in the middle that takes traffic for the website and forwards requests to the backend is very likely NGINX. Five research teams including Wiz independently discovered two memory safety vulnerabilities in NGINX's HTTP/2 handling path, both rated 9.2 under CVSS 4.0; an attacker needs no account at all to execute code on the server. F5 released fixed versions on November 18; before the upgrade is complete, temporarily disabling HTTP/2 can shut the door first.

No account, no login, and you can execute arbitrary code on the NGINX machine

Wiz's research team characterized these two vulnerabilities bluntly: unauthenticated remote code execution. In plain terms, an attacker needs no account, no login to the admin panel, to execute arbitrary code on the machine running NGINX.

Both vulnerabilities are rated Critical, with a CVSS 4.0 score of 9.2. On a 10-point scale this is the highest tier, meaning a low bar to exploit and a large impact surface. Wiz, Ikotas Labs, Metnir, Rapid7, and Xint each independently dug out the same two holes; several teams with no communication between them hitting the same spot shows this is not a fringe issue.

What can be done after successful exploitation? Wiz listed four things: execute code, steal sensitive data, tamper with the website, disrupt service. These four are not four options but a chain—first gain code execution, and the rest come along for the ride. Whatever runs on the server, the attacker can read; whatever the website returns to users, the attacker can change.

The key is where the vulnerabilities sit. Wiz is clear: the flaws are in NGINX itself, not in your application code. Even if your website program is written cleanly, with no SQL injection and no privilege escalation, as long as the NGINX layer in front is compromised, the application behind is still penetrated. Many people are used to spending their security budget on business code audits; this layer gets bypassed.

The impact also spreads laterally. NGINX is often placed at the very front of the architecture, proxying multiple applications and APIs at once; having several services on one machine is the norm. Wiz's exact words are that a single affected deployment can put multiple services at risk simultaneously. What falls is not one site but an entire row.

What kind of server gets hit?

The two vulnerabilities share only one common precondition: the affected server has HTTP/2 enabled. But beyond "enabled," each has an additional condition that determines whether your machine is certain to be hit or possibly hit.

HTTP/2 (version 2 of the web transfer protocol, faster than the old version, enabled by default on most modern sites) is the hard threshold. NGINX deployments that don't enable it are untouched by both vulnerabilities. If it's enabled, read on to match your case.

CVE-2026-123633 targets the reverse proxy layer (the server standing in front of the website to take requests and forward them to the backend). The problem lies in the handling logic when the proxy's available connection slots are scarce. Your NGINX must be forwarding traffic for a backend application and connection pressure must rise for this path to be reachable. Pure static hosting deployments that don't forward are out of its range.

CVE-2026-123007 has broader coverage. It doesn't only hit reverse proxies; NGINX serving content directly to the outside is equally affected. Wiz's test reproduced code execution over HTTPS on a 64-bit system, using an NGINX that only serves ordinary static files, has no backend application, and has no third-party modules installed. The idea that "I have no application behind it, so I should be fine" does not hold here.

In terms of versions, the impact range of CVE-2026-123007 dates back to NGINX 1.9.5 in 2015. A machine that hasn't been significantly touched since then and has kept HTTP/2 enabled has theoretically carried this problem all along. The age itself is a lead for investigation: the longer an entry point has gone without upgrading, the earlier it should be checked.

Wiz: Research (security research) official image 1
Wiz: Research (security research) official image 1

Why were these two old vulnerabilities only discovered now?

These two vulnerabilities were not newly written in. The integer underflow in CVE-2026-123007 hid in the code for over a decade, and the use-after-free in CVE-2026-123633 has existed for nearly six years. They sat there all along; no one had just dug them up before.

The reason they went unfound is specific: both are memory safety flaws. Integer underflow (a value decremented below the minimum, wrapping around to a huge number) and use-after-free (the program still reads and writes memory after it has been reclaimed) go wrong not in some single line of code but in the subtle interactions between memory operations and multiple code paths. Tracing such interactions by hand is so costly that most teams won't do it.

Wiz's research team switched tools. They used Wiz Atlas (Wiz's own AI code scanning foundation) to scan out these two vulnerabilities, then used Gemini 4 Carbon and ChatGPT 5.6 Sol to verify the impact, confirming that an unauthenticated attacker can execute code on the server. Two steps: one finds, the other confirms the finding is correct.

More noteworthy is the duplication. Multiple independent teams each dug out the same old vulnerabilities in PHP and NGINX; Wiz also states in its disclosure timeline that F5 confirmed another research team had previously reported the same CVE-2026-123633. The same batch of old problems being hit simultaneously by different people shows the threshold is not one team's luck but that the tools have collectively moved down a level.

This cuts both ways for defenders. Vulnerabilities being found earlier and fixed earlier makes software harder; but faster discovery also shortens the window from public disclosure to exploitation. As for whether AI scanning can clean out this class of flaw, there is still no third-party retest data; what Wiz published is only its own tools' results.

Wiz: Research (security research) official image 2
Wiz: Research (security research) official image 2

The patch was released on November 18; is your machine still in the gap?

Wiz reported the first vulnerability on October 21, and F5 released the security version on November 18—28 days in between. During those 28 days, attackers and you had the same code.

The timeline is clearer laid out. On October 21, Wiz reported CVE-2026-123633 to F5, attaching a confirmed working remote code execution and a remediation suggestion. F5 confirmed the same day that another research team had previously reported the same vulnerability.

On November 10, F5 gave Wiz the CVE-2026-123633 patch for review, a routine pre-disclosure check. On November 16, Wiz reported the second vulnerability, CVE-2026-123007; the next day F5 confirmed it matched the planned November 18 security release. On November 18, the security release shipped, alongside coordinated disclosure (the agreed process of publishing details after the vendor has fixed the issue).

The official recommendation is to upgrade to "the latest version containing multiple fixes," not just apply the single CVE-2026-123633 patch. The reason is practical: the November 18 release bundles fixes for both vulnerabilities at once; patching only one leaves the other in production. Is everything fine after upgrading? The next section explains.

The third step is a stopgap, and the cost must be calculated clearly. Turning off HTTP/2 means multiplexing (running multiple requests in parallel over one connection) is gone; browsers and gateways fall back to queuing by connection, and first-paint and API latency on high-concurrency sites will change. This trades performance for time; it is not a long-term solution.

There is one more thing to do right now: make the list. Which machines run affected versions, which actually have HTTP/2 enabled, which are directly reachable from the public internet. Stack these three lists together and you have the set you need to handle tonight. During the gap between patch release and production effectiveness, attackers will not wait for you.

The patch is out—is the door closed?

On the day the patch was released, production risk did not automatically drop to zero. Between getting the fix and actually having one fewer open door lie verification, coordination, and staged rollout.

Wiz puts it bluntly: turning the released fix into production protection still requires substantial manual work, from verifying the patch to coordinating rollout with service owners; none of it can be skipped. A patch is code; protection is a deployment state; between them lie human actions.

Who to find first? Wiz's advice is to first identify deployments running affected versions and configurations, and prioritize the ones attackers can reach. An NGINX that is only on the internal network and shielded by another gateway in front, and an NGINX directly attached to the public internet with HTTP/2 taking traffic, should not be handled in the same order. Exposure determines the order; the version number only determines whether to handle it at all.

Apply mitigations first during patch testing and rollout. Don't wait until the upgrade is complete to act; the server is still running old code during the test window. If you can't upgrade, disable HTTP/2 first—this is the official temporary path and the only action during this period that immediately reduces exposure.

One easy misjudgment: common memory safety mitigations like ASLR do not stop these two vulnerabilities. ASLR (address space layout randomization, which changes memory addresses on every start to make exploitation harder) is a general OS-level defense, and many people assume it raises the exploitation difficulty of memory corruption vulnerabilities. Wiz's disclosure explicitly states that exploitation of these two vulnerabilities does not rely on it as a backstop; don't treat default system configuration as a safety cushion.

Article title: Two NGINX HTTP/2 Vulnerabilities: Remote Code Execution on the Server Without Logging In

Body: Source: Wiz Research blog "Unauthenticated Remote Code Execution in NGINX: CVE-2026-123633 and CVE-2026-123007", November 18, 2026. Scope note: The vulnerability discovery and impact assessment come from joint research by Wiz with Ikotas Labs, Metnir, Rapid7, and Xint, as well as Wiz's own testing; no independent third-party retest data was seen; the fix status is based on F5's security release of the same day.

Glossary · Plain-language notes appearing in this article

The full text of in-depth pieces, research analyses, and tool tutorials, along with all podcasts, are in the members' area.

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 听说了吗
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗