CVE-2026-123007 is a CVSS 9.2 nginx HTTP/2 flaw going back to 2015. Here's how to check if you're exposed and patch to 1.31.9 or 1.30.7 today.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
If you run nginx with HTTP/2 turned on and haven't patched since November 18, 2026, you have an unauthenticated remote code execution bug sitting on your edge. CVE-2026-123007 is a memory corruption flaw in nginx's HTTP/2 flow-control code, rated 9.2 under CVSS 4.0, and it's been there since nginx shipped HTTP/2 support back in 2015. The fix is nginx 1.31.9 or 1.30.7. Here's how to check if you're exposed and how to patch it without breaking your config.
nginx2root is the name Theori's AI research tool Xint gave to a batch of five critical nginx vulnerabilities, all disclosed on the same day F5 shipped fixed releases. F5 owns nginx now, so they're the ones who coordinated the patch.
Five CVEs, all CVSS 9.2 under version 4.0 scoring:
| CVE | Component | Bug type | Needs HTTP/2? |
|---|---|---|---|
| CVE-2026-123007 | HTTP/2 flow control | Integer underflow → memory corruption | Yes |
| CVE-2026-123633 | HTTP/2 | Use-after-free | Yes |
| CVE-2026-121810 | HTTP/3 | Use-after-free in worker processes | No (needs HTTP/3) |
| CVE-2026-123634 | Proxy / SSI / mirror | Use-after-free with proxied server name variables | No |
| CVE-2026-113425 | Mail proxy | Crafted mail session RCE | No |
CVE-2026-123007 is the one getting all the headlines, and for good reason: it needs nothing more than a reachable nginx listener with HTTP/2 on. No auth, no special module, no backend app required. Researchers demonstrated code execution against nginx serving plain static files.
This part is genuinely interesting, and a little unsettling depending on your view of AI security tooling. Theori's AI system, Xint, reportedly scanned around 50,000 lines of nginx code and found four of the five flaws in about 90 minutes, without a human pointing it at specific functions first. It found the HTTP/2 underflow on October 11 and reported it through HackerOne five days later. Wiz Research independently confirmed a working RCE with an ASLR bypass for the use-after-free variant around the same time.
Two independent teams landing on the same decade-old bug within days of each other, using different methods, tells you something: this code path had been read by thousands of security engineers over eleven years and nobody caught it. Whether that's a win for AI-assisted auditing or a warning about how much latent risk is sitting in "boring" infrastructure code, I'd say it's both. I'm not ready to call this proof that AI tools are now better at finding memory bugs than humans, but I'm also not dismissing a 90-minute scan that beat 11 years of human review.
Honestly, somewhere in between, and that nuance matters for how you prioritize this.
The HN discussion on the disclosure (linked below) raised a fair point: the data requirement for one of Rapid7's exploitation paths is about 128GB of HTTP/2 traffic over a single connection. On a 10 Gbps link that's roughly two minutes. On anything under 300 Mbit/s, you're looking at over an hour, which starts bumping into connection timeouts and rate limits on a lot of real-world setups. That's not "send one packet and own the box."
But don't read that as a reason to wait. Censys found exploitation methods that need far less traffic. Public PoC code is already on GitHub. And there's a nastier wrinkle: a known Linux kernel SCTP flaw can chain with CVE-2026-123007 to escalate from an nginx worker all the way to root on the host, including out of a Docker container. If you're patching nginx, also check whether your kernel SCTP module needs attention or just blocking outright until you've patched the kernel too.
There's also a scoring curiosity worth flagging. Under CVSS 3.1, CVE-2026-123007 scores 8.1, not 9.2. The 9.2 is a CVSS 4.0 number, and the two scoring systems weigh things like attack complexity and scope differently. If your vuln management tooling still reports 3.1 scores, don't assume "only 8.1" means less urgent. It's the same bug.
Run these in order. Each one rules out a different way "I think I'm patched" turns out to be wrong.
Confirm HTTP/2 is compiled in:
nginx -V 2>&1 | grep -- '--with-http_v2_module'
Most modern builds have this by default, so expect to see output here even if you never explicitly enabled it.
Check if HTTP/2 is actually turned on in your running config:
sudo nginx -T | grep -i http2
Any uncommented listen ... http2; or http2 on; line means you're in scope for CVE-2026-123007 and CVE-2026-123633.
Verify from outside the box, not just the config file:
openssl s_client -connect yourdomain.com:443 -alpn h2 </dev/null 2>&1 | grep ALPN
If you see ALPN protocol: h2, a remote attacker can see it too.
Don't trust nginx -v for patch status if you're on a distro package. This is the trap that catches the most people. Ubuntu, Debian, and others backport the fix into their own package version while keeping a version string that still reads like the old vulnerable release. nginx -v showing 1.31.8 on Ubuntu might already be patched, and 1.31.9 on a source build is confirmed patched. Check your distro's actual advisory and changelog, not just the version number.
Check error.log for signs of probing or crashes: repeated worker process exited on signal entries around the disclosure date are worth investigating even after you patch.
nginx -s reload keeps old worker processes alive until they finish their current connections, which means vulnerable code can keep running for a while after you think you've patched. Do a full restart.Set http2 off; and remove the http2 parameter from your listen directives, then reload. This closes the door on CVE-2026-123007 and CVE-2026-123633 specifically.
It does not touch the other three bugs in HTTP/3, the proxy/SSI path, or mail proxy handling. If you run listen ... quic for HTTP/3, remove it too. If you expose nginx's mail proxy module, restrict it to trusted networks until patched. Treat the HTTP/2 toggle as triage for a maintenance window this week, not a substitute for the upgrade.
If you're behind Cloudflare or a similar CDN/WAF for your public-facing traffic, that's a layer of protection for the edge-facing exposure, but it doesn't help with internal nginx instances, load balancers between your CDN and origin, or any direct-to-origin path that bypasses it. Check your architecture, don't assume.
Patch this one even if the 128GB traffic requirement makes the worst-case exploit sound inconvenient for an attacker. Five critical bugs disclosed together, two teams independently finding working RCE on the same decade-old code, and a kernel chain to root is not a week to be lazy about nginx. I'd rather spend twenty minutes confirming my config than find out the hard way that my "it's probably fine" instinct was wrong.
The AI-discovery angle is the part I keep coming back to, though. If a 90-minute automated scan can surface bugs that survived eleven years of human eyes on one of the most audited pieces of open-source infrastructure on the planet, that's either the best argument yet for AI-assisted security review, or a reminder that "battle-tested" software has always meant "not yet broken in public," not "actually safe." Probably both, depending on the day you ask me.
If you're dealing with CVE patching as a recurring chore this quarter, we also covered the Java4Shell CVE-2026-114938 patch, another HTTP-adjacent RCE in http2shell-brotli CVE-2026-111497, and the Ubuntu kernel CVE-2026-98432 patch that's relevant if you're also dealing with the SCTP chaining risk mentioned above.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Claude Fable 5.2 agents ran unsupervised for six weeks and solved two open black hole physics problems, including a calculation stuck since 2003. Here's how the setup actually worked.
Gemini 4 Neon beats Claude Opus 5.5 on composite benchmarks (61 vs 59) at 25% lower price, but Opus 5.5 still wins long agentic coding sessions 86.9% to 79.8%.
Claude Code, Codex CLI, Pi, and OpenCode each win on different models. Runta's eval found harness choice swinging cost up to 4.4x on the same model.
Comments