NGINX HTTP/2 flaws allow unauthenticated remote code execution

NGINX HTTP/2 flaws allow unauthenticated remote code execution

Wiz Research, Ikotas Labs, Metnir, Rapid7 and Xint independently found two critical remote code execution vulnerabilities in NGINX, the software many sites use to receive and route traffic to production applications and APIs. They are tracked as CVE-2026-123633 and CVE-2026-123007. Both are rated Critical, with a CVSS 4.0 score of 9.2. Wiz published the details on November 18, 2026, the day F5, the company behind NGINX, ran its security release and a coordinated public disclosure.

Both bugs are memory-safety flaws in how NGINX handles HTTP/2 traffic, and both need HTTP/2 to be enabled on the affected server. In CVE-2026-123633, a connection's memory is freed while a request still needs it, which is a use-after-free. In CVE-2026-123007, incorrect flow-control accounting leads to an invalid memory copy. On an affected version with HTTP/2 on, an unauthenticated attacker can run arbitrary code on the server, steal data, tamper with websites or disrupt services. The flaws sit in NGINX itself, so a site can be exposed even when its own application code is secure. One NGINX deployment often fronts several applications, so the impact can reach multiple sites and APIs.

The extra conditions differ. CVE-2026-123633 affects reverse-proxy processing when available connection slots run low. CVE-2026-123007 can also hit NGINX that serves content directly: Wiz demonstrated code execution on NGINX serving ordinary static files, with no backend application and no third-party modules, in tests over HTTPS on 64-bit systems. ASLR does not protect against either flaw, because an attacker can use either one to read memory before executing code. The flow-control bug goes back to NGINX 1.9.5 in 2015, so it sat in the code for more than a decade. The use-after-free had been there for nearly six years.

The technical write-up explains the mechanics. NGINX keeps a list of connections it can close when it needs room for new ones. An HTTP/2 connection can join that list when it has no active requests, and must leave it when a request starts. If a client sends only part of a HEADERS frame, NGINX marks the connection reusable while it waits for the rest. When the rest arrives, a request stream is created and the processing counter goes up, but the connection is not removed from the reusable list. NGINX now has a live request on a connection it still thinks is safe to close. Forwarding the request to an application server opens an upstream connection; if free connection slots are low, NGINX picks reusable connections to close and can pick this one. Before closing, it tries to send a GOAWAY frame. If creating that frame fails, for example because the connection hit its internal frame limit, the error path closes the connection and frees its memory. The original request then resumes and keeps using the freed memory, which can be refilled with other data.

In the second bug, NGINX tracks the connection's remaining send allowance in a counter called h2c->send_window. When a stream is canceled while its data is still queued, the cleanup code returns the unused allowance to the counter. That refund does not check HTTP/2's maximum allowed window. If the client raised the allowance while data was waiting, the refund can push the counter past the maximum. A later bounds check in the window-update code subtracts the current allowance from the maximum; because the arithmetic is unsigned, the result wraps around to a very large positive number, so the check thinks there is plenty of room and allows further increases. The invalid counter eventually reaches the function that sends the response chain.

Wiz says its Wiz Atlas harness, which powers its AI Code scanning product, found both bugs, and that its researchers used Gemini 4 Carbon and ChatGPT 5.6 Sol to validate the impact. The disclosure timeline: on October 21, 2026 Wiz reported the use-after-free to F5 with confirmed remote code execution, an ASLR bypass and a proposed fix, and F5 said another team had reported it earlier. F5 supplied a patch for review on November 10. Wiz reported the second bug on November 16, and on November 17 F5 confirmed it matched CVE-2026-123007 and would go into the planned November 18 release. F5's release also fixes several other NGINX vulnerabilities.

Wiz recommends upgrading to a release that fixes both flaws. If that is not immediately possible, it advises temporarily disabling HTTP/2.

Key facts

  • Wiz Research, Ikotas Labs, Metnir, Rapid7 and Xint independently found two unauthenticated remote code execution bugs in NGINX's HTTP/2 handling: CVE-2026-123633 (use-after-free) and CVE-2026-123007 (flow-control accounting error).
  • Both are rated Critical with a CVSS 4.0 score of 9.2, and both require HTTP/2 to be enabled. F5 shipped the fixes in a security release on November 18, 2026.
  • Wiz showed code execution for CVE-2026-123007 on NGINX serving plain static files, with no backend or third-party modules. ASLR does not stop either flaw.
  • CVE-2026-123007 affects versions back to NGINX 1.9.5 (2015); the use-after-free had been in the code for nearly six years.
  • If upgrading is not immediately possible, Wiz advises temporarily disabling HTTP/2 on affected instances.

Why it matters

NGINX often sits at the edge and handles traffic for several applications, so one vulnerable deployment can put many sites and APIs at risk. These flaws are in NGINX itself, which means secure application code does not protect you. One of them had been in the code since NGINX 1.9.5 in 2015. Wiz also argues the discovery points to a wider shift: memory-safety bugs that were hard to find because they need tracing subtle interactions across code paths are now within reach of AI models, for researchers and attackers alike. It links this to the PHP2Shell bugs it reported last month and to Java4Shell, which Calif disclosed the day before.

Who it affects

Anyone running NGINX with HTTP/2 enabled on an affected version. CVE-2026-123633 applies to reverse-proxy processing when available connection slots run low. CVE-2026-123007 can also affect NGINX that serves content directly, and its oldest affected version is NGINX 1.9.5. Because one NGINX instance can front several applications, the exposure can span multiple websites and APIs behind it. Security teams carry the practical load: Wiz notes that turning a published fix into protection in production still takes substantial manual work, from validating patches to coordinating rollouts with service owners.

How to use it

Identify NGINX deployments with HTTP/2 enabled and upgrade to a release that fixes both vulnerabilities. Prioritise deployments an attacker can reach. If you cannot upgrade right away, temporarily disable HTTP/2 on affected instances, test application compatibility, reload NGINX and confirm the old workers have exited. Re-enable HTTP/2 after upgrading. Investigate unexpected NGINX worker crashes or commands launched by NGINX, but Wiz stresses these are general warning signs, not confirmed indicators of these vulnerabilities.

How solid is it

The vulnerabilities were found independently by five teams and disclosed in coordination with F5, which supplied a patch and ran the security release on November 18, 2026. The analysis comes from Wiz, one of the finders. Wiz says it confirmed remote code execution for both bugs and demonstrated it for CVE-2026-123007 on static-file serving over HTTPS on 64-bit systems. The claim that AI is dissolving the barrier to finding such flaws is the post's own opinion, from a company whose AI scanning product found the bugs. The source reports no in-the-wild exploitation and mentions no public exploit code.

Risks and caveats

Standard memory-safety defences such as ASLR do not help, since either bug gives an attacker memory read access before code execution. Wiz warns that faster discovery and a shorter gap between disclosure and exploitation add to security teams' workload. The disclosure text names no fixed NGINX version numbers, and the exact configuration change for disabling HTTP/2 is not given here, so check F5's release notes. Disabling HTTP/2 is a temporary measure that needs compatibility testing. The source does not say how many NGINX servers are exposed.

“Complexity can no longer be relied upon as a barrier to discovery or exploitation.”

— Wiz Research