Nginx fixes two critical remote code execution flaws in HTTP/2

- Reading
- 5 min
- Tables
- 1
Researchers found two critical remote code execution vulnerabilities in Nginx, the world's most widely used web server, and maker F5 has already released mandatory fixes for anyone using HTTP/2.
Two flaws that compromise Nginx without needing a password
On November 18, 2026, F5, the company that maintains Nginx, published a security fix package that includes two critical unauthenticated remote code execution (RCE) vulnerabilities: CVE-2026-123633 and CVE-2026-123007. Both received a 9.2 score on the CVSS 4.0 scale, the ceiling of the critical category, according to the Wiz Research report.
The two flaws were independently discovered by five different research teams: Wiz Research, Ikotas Labs, Metnir, Rapid7, and Xint. All of them reported the issue to F5, which coordinated the disclosure. The timeline shows how tight it was: Wiz reported the first bug (CVE-2026-123633) on October 21, received a patch for review on November 10, reported the second one (CVE-2026-123007) on November 16, and saw both fixed together two days later.
The most striking part is the age of the flaws. CVE-2026-123007 has existed since version 1.9.5, released in 2015: more than a decade of vulnerable code in production. CVE-2026-123633 has been present since version 1.19.7, from about six years ago. Neither is a recently introduced bug; they are memory defects that simply had never been found before.
How the attack works in practice
Both require HTTP/2 to be enabled on the server, but the exploitation paths differ. In CVE-2026-123633, an HTTP/2 connection that is still processing a request can remain marked as "reusable" inside Nginx. When the server needs to free up memory and closes that connection by mistake, the code that was still processing the request keeps using a pointer to already-freed memory: a classic use-after-free, which Wiz turned into code execution with an ASLR bypass included.
CVE-2026-123007 is a flow-control issue. Nginx keeps a counter (h2c->send_window) that tracks how much data it can still send to the client. When a request is canceled with data already queued, Nginx returns that quota to the counter without checking the protocol's maximum limit. The counter overflows, a bounds check that uses unsigned arithmetic suffers an underflow, and the invalid value reaches the routine that copies data to the TLS connection, opening an out-of-bounds memory copy.
The detail that matters most in practice: Wiz demonstrated RCE in CVE-2026-123007 while serving only static files, with no application behind it and no third-party module. This means that even an Nginx instance configured only to serve HTML and assets, with no backend at all, can be compromised.
The fix package includes eight more flaws
F5's security release didn't fix only these two. The bulletin lists eight other vulnerabilities, with severities ranging from medium to critical:
| CVE | Description | Impact | CVSS |
|---|---|---|---|
| CVE-2026-121810 | Use-after-free in HTTP/3 | RCE when configured as an HTTP/3 proxy | 9.2 |
| CVE-2026-123634 | Use-after-free in the DNS resolver | RCE when the attacker controls the upstream hostname | 9.2 |
| CVE-2026-113425 | Use-after-free in the mail proxy module | RCE on an email server with auth_http | 9.2 |
| CVE-2026-118899 | Infinite loop when reading the request body | Denial of service with TLS 1.3 | 8.7 |
| CVE-2026-125283 | Buffer overflow with alias and try_files | Theoretical RCE in an unusual configuration | 6.9 |
If your installation uses HTTP/3, the mail proxy module, or dynamically resolves DNS from upstream hostnames, it's worth checking each of these rows, not just the two headline flaws.
What changes for those running infrastructure in Brazil
Nginx is the invisible foundation of much of what runs in production here: from a simple VPS serving a Node.js app behind a reverse proxy to Kubernetes cluster ingress controllers and load balancers in front of APIs. Wiz estimates the software serves more than 225 million sites worldwide.
The central point of the report is that these flaws are in Nginx itself, not in the application behind it. A Node.js API, a Django backend, or a Next.js frontend with flawless code still remain exposed if the Nginx in front of them has HTTP/2 enabled and is on a vulnerable version. This takes the patching decision out of the product team's hands and puts it on the infrastructure team's table, with a short deadline.
Since a single Nginx instance usually serves several applications at once, the blast radius of a successful exploit tends to be larger than that of an isolated flaw in a single application.
How to mitigate now
F5 fixed both flaws in the following versions:
- Nginx OSS: 1.30.7 (stable) and 1.31.9 (mainline) fix both CVEs.
- Nginx Plus: release R33 fixes both; specific builds 37.0.7.1/37.2.0.1 cover CVE-2026-123633 and 37.0.8.1/37.3.0.1 cover CVE-2026-123007.
When an immediate update isn't feasible, Wiz's recommendation is to temporarily disable HTTP/2:
- On instances running Nginx 1.25.1 or newer, set
http2 off;in the affectedserverblocks. - Remove the
http2parameter from anylistendirective. - Test application compatibility and reload Nginx, confirming that the old workers have terminated.
- Re-enable HTTP/2 only after applying the definitive update.
As a general warning sign (not confirmation of an attack), it's worth investigating unexpected Nginx worker crashes or commands spawned by the process itself.
AI is changing who finds these bugs first
One technical detail of the report is of interest to anyone working in applied security: Wiz used Wiz Atlas, an AI-based code-scanning engine, and validated the impact of both flaws with Gemini 4 Carbon and ChatGPT 5.6 Sol before confirming code execution.
The fact that five different teams independently found the same bugs, after more than a decade hidden in the code, is no coincidence according to the researchers themselves:
The flaws were already there; AI helped us find them.
Wiz Research team
Wiz's argument is that AI models increasingly better at tracking subtle interactions between memory operations are tearing down the practical barrier that used to protect complex memory bugs, such as use-after-free and integer underflow, from being discovered. The same pattern had already appeared in recent flaws in PHP (nicknamed PHP2Shell) and OpenJDK (Java4Shell, disclosed on November 17, 2026).
What remains open is the other side of that coin: if AI tools speed up discovery by researchers who report responsibly, the same capability can speed up discovery by those who want to exploit before the patch. For infrastructure teams, this reinforces a practical and unglamorous point: the window between disclosure and exploitation tends to shrink, so the speed of applying the patch in production matters as much as the quality of the patch itself.
Translated from the Brazilian Portuguese original · Read the original
Ruby flaw exposes default Rails apps to memory corruption; RCE not demonstrated
An advisory from Veria Labs describes a one-byte heap overflow in the Ruby interpreter, reachable from Rails' default configuration and present in GitLab, GitHub Enterprise, Mastodon, Discourse, and Chatwoot, but the company has not demonstrated remote code execution.





