- Rapid7 and F5 disclosed CVE-2026-123007, an nginx HTTP/2 flaw that can allow unauthenticated remote code execution under certain conditions.
- Rapid7's exploit route needs about 128 GB of HTTP/2 data and a specific CPU feature path. Rapid7 does not present that volume as a floor for every exploit.
- Teams should confirm the fixed nginx version, check where HTTP/2 terminates, and not treat the exploit's limits as a substitute for patching.
A bug in a rule the protocol already states
Some vulnerabilities come from hard problems. This one comes from a rule that the HTTP/2 standard states clearly and that one piece of code did not enforce. The lesson for leaders is that a flaw this deep can still be limited by something as plain as bandwidth.
Rapid7 and F5 disclosed CVE-2026-123007 on November 18. Rapid7 describes it as an integer underflow in nginx's HTTP/2 handling. It can allow unauthenticated remote code execution against affected servers "under certain conditions". Rapid7 researcher Stephen Fewer found it.
How the flaw works
HTTP/2 lets one connection carry many requests at once. The client tells the server how much response data it may send, a budget called flow-control credit. nginx tracks that budget in counters. The standard caps the connection counter at 2,147,483,647 bytes.
Rapid7 says nginx stores that counter in an unsigned number that can hold far more than the cap. So nginx has to enforce the limit in its own logic. When a request is cancelled, nginx hands back the credit it had reserved. Rapid7 found that this step does not check whether the counter now exceeds the cap.
A second behaviour keeps the connection alive. When a settings change pushes a stream past the cap, the standard, RFC 9113, requires the connection to end with an error. Rapid7 says nginx closes only the affected streams and carries on.
The next credit update then subtracts the cap from a counter that is already above it. The result cannot be negative, so it wraps around to a huge positive number. The attacker can then raise the counter to an enormous value.
Later, nginx reads that oversized counter as a negative response length. Rapid7 says this creates an invalid buffer and leads to a memory copy with a huge byte count. The result is out-of-bounds reads and writes. With a suitable memory layout, Rapid7 says an attacker can disclose data, interrupt service or run code with the nginx worker process's permissions.
The exploit takes a lot of traffic
Rapid7 built a working proof of concept against a Docker image of nginx 1.31.8 with HTTP/2 and TLS enabled. Its route needs approximately 128 GB of HTTP/2 data. Each credit update is only 13 bytes, so the attacker must send billions of them.
The 128 GB belongs to one specific path. Rapid7 does not offer it as the least any attacker could ever need. Other exploit strategies may be viable, though none are currently known.
Rapid7 gives the timings. Its demonstration took two and a half minutes. Over a 10 Gbit/s network the transfer takes about two minutes, and over 1 Gbit/s about seventeen minutes. nginx's default connection lifetime is one hour, so a link slower than roughly 300 Mbit/s cannot finish in time.
Two more limits apply. If a proxy sits in front of nginx, the attacker's exact HTTP/2 frames will likely not be passed on in a way that works. And Rapid7's route depends on a specific x86-64 glibc memory-copy routine, chosen at runtime from the avx512f and avx512vl CPU features.
What this changes for a security budget
These limits are real, and they are Rapid7's own. They are not a reason to delay patching. They do change the questions a leader should ask.
The first is about exposure. Rapid7 says a proxied HTTP/2 connection will likely not carry the attack frames intact. Ask your team where HTTP/2 actually ends. If nginx faces the internet directly with HTTP/2 on, that matches the network condition Rapid7 tested. The exploit also depends on the CPU and the glibc memory-copy path, so direct exposure is only part of the picture. If a load balancer or CDN terminates HTTP/2 first, the picture differs. Your team should confirm this from your own configuration.
The second is about detection. Rapid7's route means on the order of 128 GB from one client on one connection, lasting minutes. Does your monitoring alert on a single very heavy connection?
The third is about the fix. The Rapid7 headline marks the issue as fixed. The excerpt we reviewed does not list patched version numbers, so take them from F5's advisory. Then ask which of your servers, container images and appliances bundle nginx, and who owns each one.
What to ask your team this week
Ask for an inventory of internet-facing nginx instances with HTTP/2 enabled, and the version each one runs. Ask which sit behind a proxy that terminates HTTP/2. Ask whether network monitoring can flag a single connection moving unusually large volumes. Ask who confirmed the fixed version.
The cost of Rapid7's exploit is measured in bandwidth and time, and Rapid7 notes other exploit strategies may exist, though none are known. A bug that needs 128 GB of traffic is still fixed by a patch, not by hoping the attacker's connection is slow.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Rapid7.





