BLACKOUT MAP

Security ·

NGINX2Root: a 2015 bug in nginx's HTTP/2 code lets strangers run code, and exploits are public

Este artículo está en inglés.

Compartir
Updated 19 November, 13:30 UTC
  • First version.

nginx runs a large part of the web, and on 18 November it got its biggest security fix in years. The worst of five bugs sits in the code that speaks HTTP/2, has been there since 2015, and can let a stranger run their own code on the server without logging in. Code to attack it is already public.

Status on 19 November 2026, 13:30 UTC

The fix is out: nginx 1.30.7 (stable) and 1.31.9 (mainline), NGINX Plus 37.0.8.1 and 37.3.0.1. Affected: 1.9.5 to 1.31.8 with HTTP/2 switched on. After updating, restart nginx: a reload keeps the old program running.

Debian and Ubuntu: no fixed package yet for their stable releases. Until then, switch HTTP/2 off.

Attacks: none reported, and none of the five is on CISA's list of exploited flaws. Exploit code is public.

The bug in one minute

HTTP/2 lets a browser send many requests over one connection at once. To stop either side from being flooded, both keep a counter of how much data the other may still send, called the flow-control window. Small control messages raise or reset that counter.

In nginx, carefully chosen messages can push the counter past the limit it is supposed to stay under. Xint, who found the bug, describes it as an integer overflow: the number wraps around, and nginx then reads and writes memory outside the buffer it meant to use. Most of the time the worker process simply crashes. With more work, an attacker can turn it into running their own commands, with the rights of that nginx worker.

Two details make it nasty. It needs nothing special on the server: Wiz says it got code running on an nginx that only served ordinary static files. And the attack is made of control messages (WINDOW_UPDATE and RST_STREAM), not requests, so it doesn't show up in the normal access log.

Five bugs, one name

"NGINX2Root" is the name Xint gave the set. F5, the company behind nginx, rates all five 9.2 out of 10 (CVSS 4.0). Whether one matters to you depends entirely on your configuration:

Need HTTP/2

  • CVE-2026-123007: the flow-control bug above. Any server with HTTP/2 on, 1.9.5 to 1.31.8.
  • CVE-2026-123633: memory used after it was freed, when nginx passes HTTP/2 requests on to a backend and runs short of connections. 1.19.7 and up.

Need something rarer

  • CVE-2026-121810: HTTP/3 (QUIC) with a backend. 1.25.0 and up.
  • CVE-2026-123634: the DNS resolver, with a backend name built from variables plus a mirror or SSI subrequest.
  • CVE-2026-113425: nginx as a mail proxy with auth_http. No valid login needed.

Turning HTTP/2 off blocks the first two only. If you use HTTP/3, the resolver or the mail proxy, only the update helps.

Found by five teams

F5 credits Ikotas Labs, Metnir, Rapid7, Wiz and Xint, who reported the bugs separately. Xint published its timeline:

  • 11 October An Xint researcher finds the HTTP/2 bug while preparing for the ZeroDay Cloud contest. A few days later an automated Xint scan finds the other four.
  • 16 October Reported to F5 through HackerOne.
  • 27 October The reports reach the nginx developers.
  • 10 November Xint confirms the fixes work.
  • 18 November nginx 1.30.7 and 1.31.9 are released, with fixes for ten flaws in total.

The bug itself is older than most of today's web: it came with nginx's very first HTTP/2 release, 1.9.5, in 2015.

How many servers?

Censys counted about 9 million nginx services that show an affected version and have HTTP/2 on, and another 103 million that don't show their version at all. Xint's own search found 56.7 million nginx hosts, of which 11.94 million showed a version, and 98.9% of those were in the affected range.

Both counts overstate it in one way: Linux distributions fix bugs without changing the version number, so an "old" version can be a patched one. And understate it in another: most servers hide their version.

Is there an exploit?

Yes, more than one, with very different claims:

  • Rapid7 published working code that leaks memory addresses to get around ASLR (the protection that hides where things are in memory). It says it needs about 128 GB of HTTP/2 traffic and two minutes to an hour, and that on a slow line nginx's own one-hour connection limit stops it.
  • Censys says it reproduced the bug in its lab without that much traffic, and published its own code.
  • Xint and Wiz say their exploits work reliably, including over the internet on AWS, Google Cloud and Azure. They are keeping the final steps private.
  • F5 says code execution is only possible when ASLR is off or bypassed. Wiz and Rapid7 both say they bypassed it.

So the public code is slow and noisy, but the people who know the bug best say it can be done properly. Expect the gap to close.

From nginx to root

An attacker who gets in through nginx runs as the nginx worker, usually a user with few rights. Xint went one step further: it chained the bug with a separate flaw in the Linux kernel's SCTP code (present since Linux 4.15 in 2018) to become root, even escaping from a Docker container to the host. That kernel flaw had no CVE and no fix yet when this was written. Servers that don't use SCTP can stop the module from loading with a line install sctp /bin/false in a file under /etc/modprobe.d/ (this doesn't help if it is already loaded or built in).

Are you affected?

You need both

  • nginx 1.9.5 to 1.31.8 without your distribution's fix, NGINX Plus before 37.0.8.1 / 37.3.0.1, or an older R-series release.
  • HTTP/2 on: http2 on; in a server block, or http2 in a listen line.

You're safe from the worst one with

  • nginx 1.30.7 or 1.31.9, or a package with the fix.
  • HTTP/1.1 only.
  • Apache, Caddy, LiteSpeed, IIS: different software (Apache had its own HTTP/2 bug last week).

Look out for nginx you didn't install yourself: Kubernetes ingress controllers, Docker images, control panels and appliances often bundle it. F5's own ingress controller (5.8.0) and Gateway Fabric (2.9.0) are fixed. The community ingress-nginx project was archived in March and gets no fixes.

What to do

  1. Find out if you run it. nginx -v shows the version, but on Debian or Ubuntu check the full package version with dpkg -l nginx and compare it with your distribution's page for CVE-2026-123007. nginx -T 2>/dev/null | grep -n http2 shows where HTTP/2 is on.
  2. Update and restart. Take 1.30.7 / 1.31.9 or your distribution's fixed package, then systemctl restart nginx. Not reload: a reload keeps the old program in memory.
  3. No fixed package yet? On 19 November Debian 12 and 13 were still listed as vulnerable (only Debian's development branch had the fix), and Ubuntu had no fix either. Then switch HTTP/2 off: change http2 on; to http2 off; in every server block (or remove http2 from listen lines on nginx before 1.25.1), run nginx -t, and restart. Sites keep working over HTTP/1.1. gRPC needs HTTP/2, so test that first.
  4. Don't count on a CDN. Cloudflare and similar services keep talking HTTP/2 to your visitors even after you switch it off, so nobody notices. But they only protect you if your server can't be reached directly. Test from outside with openssl s_client -connect YOUR-IP:443 -servername YOUR-SITE -alpn h2: if it answers ALPN protocol: h2, HTTP/2 is still reachable.
  5. Look for signs. Xint suggests watching for one long connection from a single address that keeps receiving data while sending little, workers crashing in error.log, and any program started by an nginx worker. That last one is the clearest sign of a break-in.

We did the same on our own servers: HTTP/2 is off until our distribution ships the fix. Visitors still get HTTP/2 from Cloudflare in front of us.

The other five fixes

The same releases fix five more flaws. One deserves attention: CVE-2026-118899, rated major by nginx, lets a client send nginx into an endless loop while it reads a request body over TLS 1.3, in 1.13.0 to 1.31.8. It doesn't need HTTP/2, so switching HTTP/2 off doesn't help against it. The others are rated minor or medium: a crash with PROXY protocol v2 (1.31.4 to 1.31.8 only), an overflow with alias and try_files, a client certificate check that can be skipped through SNI in the stream module, and weak checks in ssl_preread.

Why this is on Blackout Map

This is a software flaw, not an outage, and it won't show on our map. But its most common ending is a crash, and nginx serves so much of the web that crashing it at scale would be an outage for millions of sites. Our Security watch tracks the flaws that are really being used. Read also Can a software flaw take the internet down?

Sources

Comparte este artículo

Blackout Map is a live map of internet outages and shutdowns around the world. Open the map · Security watch · RSS · Follow us on X: @Blackoutmap_net

Todos los artículos

Sigue leyendo

Security · 12 Nov 2026 146Apache's HTTP/2 flaw: found twice in eight days, fixed in 2.4.71CVE-2026-119618 is a race in Apache's mod_http2 that can let an attacker run code without logging in. Only servers with HTTP/2 on are affected. Who found it, who is at risk, the dispute over an exploit, and what to do. Security · 6 Nov 2026 359How one asterisk opened a door in OpenSSH: the *sshole bug, explainedCVE-2026-123158 lets a 257-character password into accounts that have no password, on Alpine and other musl systems. How three sensible fixes built it, who is really affected, what some reports get wrong, and what to do. Explainer · 21 Oct 2026 239Can a software flaw take the internet down?Ten years after the Mirai attack on Dyn: how security flaws turn into outages, the times they did, and what Blackout Map can and cannot see.

Más del blog

Security · 17 Nov 2026Why hackers will never wipe out your debtOn 16 November malware ran on the websites of HSBC, Equifax and 176 other organisations, and the internet joked it got "everything but credit card debt and student loans". Why that is no accident: what the attack really did, and how banks keep copies of what you owe out of any hacker's reach. Security · 15 Nov 2026Scrapers that never use the same address twice, and how we stop them without CAPTCHAsResidential proxy networks send each request through a different person's home internet, so blocking addresses is useless. How we spot them by network instead, what a one-second JavaScript check does, and what it still can't stop. Outage · 10 Nov 2026Internet partial outage in Central African Republic: what we measuredBlackout Map measured an internet partial outage in Central African Republic from 20:45 UTC on 7 November 2026. What the data shows, and how to follow it live. Analysis · 6 Nov 2026 134Is 72 hours of supplies enough? What 93 nationwide power outages showCanada and the UK tell people to be ready for three days without power. We measured 93 nationwide outages since 2022: nine in ten ended within three days. The eight that didn't were all on islands, and local storm damage can take weeks to repair. Fact check · 5 Nov 2026 1,987Every "global blackout" date so far, and what really happenedA running list of viral global blackout dates: what was claimed, what was real, and why it can't happen that way. Includes 13, 17 and 18 November 2026, with the live status of the internet. Analysis · 5 Nov 2026Election shutdown watch: where the internet could be cut nextThe next elections to watch for internet shutdowns, starting with Cabo Verde, The Gambia, Guinea-Bissau and South Sudan, with each country's live internet status and what happened at this year's votes. Security · 3 Nov 2026 116Ivanti Connect Secure: three exploited flaws, and the first fix wasn't enoughAttackers chained three flaws to get root on fully updated Ivanti Connect Secure VPN appliances. All three are now on CISA's list of exploited vulnerabilities. Which versions fix it, what the attackers did, and what to check. Fact check · 31 Oct 2026 9,087¿Habrá un apagón mundial el 17 de noviembre? Lo que dicen los datosCirculan avisos de un apagón mundial el 17 de noviembre: sin luz, sin internet durante tres meses y las cuentas a cero. De dónde sale, por qué no puede ocurrir así, qué hay previsto de verdad ese día y el estado en directo de cada país. Analysis · 24 Oct 2026 144LaLiga's IP blocks: why half a million websites break in Spain on match daysTo fight pirate football streams, Spain's ISPs block shared IP addresses during LaLiga matches, and hundreds of thousands of unrelated websites go down with them. What OONI found, what we measured, and what you can do before Barcelona – Real Madrid. Security · 21 Oct 2026 999Node.js security release: ten fixes, one that can run codeOn the same day as PHP2Shell, Node.js fixed ten flaws in its 22, 24 and 26 lines, including a Brotli decoder bug that can allow remote code execution and several ways around TLS hostname checks. What is affected and what to do. Security · 21 Oct 2026 16,569PHP2Shell: what we know about the two critical PHP flaws (now exploited)Two critical flaws in PHP, the language behind about 70% of websites, can let anyone run code on a server with one request. The fix is out and attacks are confirmed: how the attack works, what is confirmed, what to do, and what it could mean for the internet. Retrospective · Iran, February 2026 692Iran's wartime blackout: 88 days cut off from the worldFrom the first strikes on 28 February to 26 May 2026, Iran was almost completely cut off from the internet: 2,093 hours, the longest nationwide shutdown on record. What happened, how it worked and what it cost. Retrospective · West Africa, March 2024 180West Africa lost four undersea cables in one dayA suspected underwater landslide off Côte d'Ivoire broke four cables at once in March 2024. Ghana was in blackout for almost a week, Côte d'Ivoire fell to 1% of normal. Retrospective · Canada, July 2022 176How a maintenance update took down Canada's RogersIn July 2022 one update in Rogers' core network cut mobile, home internet, debit payments and emergency calls for millions of Canadians for most of a day. Retrospective · Tonga, January 2022 161Tonga: the volcano that cut a country offThe 2022 Hunga Tonga eruption shredded the only cable linking Tonga to the world. The country was almost completely offline for five weeks. Retrospective · Worldwide, October 2021 225The six hours Facebook disappeared from the internetIn October 2021 one routine command took Facebook, Instagram and WhatsApp off the internet for billions of people. No attack, no country outage. Retrospective · Myanmar, February 2021 177Myanmar 2021: a coup, and an internet curfewAfter the February 2021 coup, Myanmar's military cut the internet nationwide, then every night, then all mobile data: a shutdown shaped by the hour. Retrospective · Tigray, Ethiopia, November 2020 171Tigray: two years without a connectionFrom November 2020, the war in Ethiopia's Tigray region came with a phone and internet blackout that lasted about two years. Retrospective · Iran, November 2019 175Iran, November 2019: a country offline for a weekDuring the 2019 fuel protests Iran cut itself off from the internet almost completely, while keeping parts of its domestic network running. Retrospective · Jammu and Kashmir, India, August 2019 172Kashmir: 550 days without full internetThe shutdown that began in August 2019 became the longest ever imposed in a democracy. High-speed mobile internet only returned in February 2021. Retrospective · Egypt, January 2011 163Egypt 2011: the night a country switched off its internetIn January 2011, two days into mass protests, Egypt withdrew almost all its routes from the internet within minutes. It stayed dark for five days. Retrospective · Pakistan, February 2008 197The day Pakistan took YouTube offline for the whole worldIn 2008 a national block on YouTube leaked into the internet's routing system and made the site unreachable worldwide for about two hours.