Well, whoever picked the name HTTP2Shell certainly didn't use any intelligence, artificial or genuine.
It's so meaningless, broad and undescriptive of the actual issue. No it probably doesn't affect an HTTP server with nothing else behind it, so why is HTTP being blasted?
I think we've all seen where an LLM names a session, issue, or PR based on its early understanding and fails update it later.
AI isn't being asked to review the title, and it doesn't have experience or think, so of course it's not going to flag it.
No human reviews anything anymore so of course they aren't going to catch it.
Edit: AI for as good as it is at "writing" is terrible at writing and particularly at naming things. I had an AI name a client OpenAICompatibleEmbeddingClient. "Brooooo.. I think "compatible" is implied, who names stuff like that?!"
Much prefer the Broccoli2Shell name & site (for the same vuln): https://broccoli2shell.com/. For starters, I'm pretty sure you can hit this without HTTP if you do any manual Brotli decoding, HTTP just happens to be the most common case that comes up.
To me, the surprising part is how many of these implement brotli compression of request (not just response) bodies, given how obscure a feature that is.
In any case all this sensationalism is starting to get annoying. Patch your stuff and move on.
Edit: actually, between all the sensationalism, its hard to tell from the article how many of these actually support brotli compressed request bodies in common configs vs some other use of brotli (which is bad but nowhere near as critical)
It seems like node supports it with body parser enabled.
But im having trouble finding any docs suggesting anything else supports it on inbound requests.
Obviously there are other scenarios where it is exploitable, but it is a much less severe vuln if the attacker must first force the program to make an outbound request to a server of its chosing rather than just connect to a vulnerable server.
I wonder if we'll see more of this thanks to AI. Extremely complex features that nobody uses and that were easy to add, which incidentally tremendously expand the attack surface. One would hope that maintainer would be wise to avoid that, but for how long?
A few month ago, everyone was advising to use cool-downs in your package upgrades. Now we're constantly overriding those because security incidents will be exploited immediately.
Guess that's just how it's gonna be from now on.
Article claims OpenAI found this in February, but I don’t think that’s correct. I think Google made a change to the Brotli source code in February that fixes the issue, but hadn’t realized it was a security issue and therefore hadn’t cut a release.
Let’s say I have a node.js app behind an Amazon elastic load balancer which terminates TLS. My understanding is that in nearly all configurations, my app would be unaffected because the ELB would handle decompression. Is that generally correct?
I don’t know about that specific software, but I would be surprised if it decoded message bodies, which is entirely unrelated to TLS termination and shouldd be entirely unrelated to routing. So I would expect you would still be vulnerable.
I had to check all of our Python deployments and installation BOMs and was quite relieved - turns out that none actually installed brotli libraries. HTTP libraries like httpx, requests and frameworks consider them optional and autodetect their presence.
An example of the cost of added complexity I suppose. Brotli is much more complicated than the gzip/deflate it replaces. ... and of course we still have to expose gzip to attack as well.
Every time you add something you lose something...
And strictly speaking you cannot trust the pip upgrade of pip with the vulnerable cert in the trust store due to certifi, which you also cannot upgrade safely with pip for the same reason. All of these are at one out of two compromises from being MITM'd during the upgrade.
So strictly we have to burn everything python to the ground and reinstall entirely.
Heh. This is part of why I always preferred getting my python runtime and all these core libs from my linux distro packager. But python die-hards look at me like I'm insane when I suggest this.
Even in today's LLM-addled world, I still think it's a better tradeoff for smaller system/site operators. Get the distro-level testing and regular auto-patching, rather than telling yourself it is a good idea to "freeze" your python world at the pace of your slowest app/system integrator.
It's the same reason I think docker is an anti-pattern for these small groups. You really cannot safely use these strategies to bake an image and divorce yourself from the maintained distro layer unless you are going to adopt CI/CD that does the same thing. Keep tracking all these patches to the entire runtime with all the dependencies, rebuild, re-test, and redeploy.
It's total cargo culting to adopt the freeze/image techniques and just abandon apps in place as if that is robust!
But, it's also more than just automatically updating everything. There is an editorial aspect to distro management to help secure the supply chain to your app. Most smaller orgs cannot feasibly do this as well as the regular linux distro maintainers. Some large orgs can probably do it better, but that doesn't matter if you aren't doing it large org style with dedicated security folks tracking all this against well-developed policies and procedures...
Or - you can just upgrade pip with a SHA hash pin, or set the magic env vars to use the system bundle.
You do need to be careful about creating new venvs with the old version of pip bundled in ensurepip in older Python releases. `uv` kindly patches that to newest pip, but not every Python binary does.
Quit being overly pedantic about how contrived the example is and struggle with the underlying fact that the compression type is negotiated between the client and server. If the compression happens on the HTTP request (as some of the comments here suggest), then the server must use that compression method or it fails to read the body of the request.
Lots of apps will, for example, bring in your Google avatar (i.e. they fetch its URL for ingestion) as a default avatar if you use them for SSO. Slack included, IIRC.
Honestly it's time to get rid of C as a language used where security is a factor. Enough already. Why the hell is a lib like brotli that processes public general internet traffic, and maintained by google, not in rust?
Security is a factor everywhere. And C is everywhere. There's not going to be a "rewrite everything in Rust" project until after there is some legit penalty for selling / offering software in C that could be more secure in another language.
I think Mozilla got it right when they abandoned the effort to rewrite the entire browser in Rust and instead focused on analysis of which modules needed priority rewrites. If a module has historically had zero memory-related bugs, then there are other higher priority things to spend time on.
That said, Google Project Zero should have had Brotli on their radar. Hopefully they still have people and resources.
git for windows is vulnerable even in its latest release (2.56.0.windows.3), careful pulling repos over https from random ass servers (prolly uncommon but hey). they already merged the fix on the repo, it's just in a snapshot build only; should prolly arrive on Monday (2.56.0.windows.4).
also affects codex (gui), which ships its own git, could be more of a problem there, when the LLMs get quirky at night. still unlikely to be a problem though.
speaking of, update codex (and wipe its cache at ~/.cache/codex-runtimes), ships its own node. update python.
stuff like wireshark and imhex remain vulnerable, even on their latest repo commits, which may or may not be spicy.
also Microsoft 365 Copilot still has its own unpatched brotlidec.dll from Oct 7, if anyone uses that. no fix available for that one afaik, not sure about how it can be made to access the vulnerable decoder.
Because a local software vuln is much less scary, less exploitable, and more common than an exploit like PHP2Shell, which requires basically no effort to exploit a large chunk of publicly exposed webservers. Basically, local vulns have been a thing forever and GP is off topic.
I really don't understand how so many different things are affected by the same brotli bug/exploit.
I even read the patch for the C code (patch which, btw, seems a bit hacky to me with isSafe? methods added and whatnots but YMMV: now I understand things had to be patched in a hurry but this patch really looks like "lipstick on a buffer overflowing C pig" to me).
What's the link between Python / Node / C++ / libbrotli / Envoy / Bun: are those some all using the same C brotli encoder/decoder?
Can someone who knows more explain why all those are vulnerable?
> As for unaffected software, Apache’s standard mod_brotli does not decode Brotli
Because this one only encodes and doesn't decode and the bug is in the decoder? (for I take it mod_brotli, with such a name, still does something brotli related and isn't, say, a TOTP token generator nor a SAT solver?).
I mean overall, sheesh: could we get better explanation / better writeups when such big fuckups happen?
Yes. All these language runtimes compile against common libraries. Compression libs are one of the most common common libraries that platforms like node, ruby, dotnet, python, and etc need to compile.
Google provides the Brotli implementation. Package maintainers slurp that up and create ie libbrotli-dev. When you compile node it uses that for brotli support.
Brotli was a library released by the Google team that developed it. There weren't lots of independent implementations; they were all from the same source code.
As I've mentioned numerous times at this point, relying on third party code is a massive liability (supply chain attacks/sloppy code). My hand rolled brotli implementation is not only faster than libbrotli but also immune to this vulnerability (yes, I checked). Roll your own code people.
Why are we still so reliant on HTTP? HTTP came out in 1990 (36 years ago) and the most common version of HTTP still in use, 1.1 (RFC2616), came out in 1999 (27 years ago)?
Yes, I know HTTP 2 and 3 and SSE are attempts to make it faster and more parallel, but its still slow. Its slow because its unidirectional with a full round trip and by default its a new socket per each message. Yes, I know the real reason HTTP became and remains popular is because its sessionless... in theory. Still, there are better options available. You can do so much more with a full duplex channel.
HTTP 1 is full of parser differential vulnerabilities, which 2 fixes. That is probably its biggest flaw.
SMTP is too chatty - it uses too many round trips for efficient transfer across the internet. It can take several seconds just to get to the point of sending the email data. It also duplicates information between the SMTP envelope and the email body which creates vulnerabilities when they don't match.
For APIs that don't have to work in a browser I suggest JSON lines over TCP, or over TLS.
Agreed. The issue with all these archaic standards is that they were rolled decades ago back when certain practices were either outright not being employed OR that hardware wasn't fast enough. Ironically, HTTP itself is a literal nightmare to parse (performance wise) on top of having like... ten? distinct vulnerability classes. The world desperately needs new standards.
Just open a TCP socket and send JSON messages. Or if you want a bit more efficiency, fixed-format binary messages, but remember to include a version number.
For custom server implementations, agreed. But the web still depends on HTTP and we sadly need the major infra providers to agree on a new standard otherwise HTTP isn't going anywhere
Well, whoever picked the name HTTP2Shell certainly didn't use any intelligence, artificial or genuine.
It's so meaningless, broad and undescriptive of the actual issue. No it probably doesn't affect an HTTP server with nothing else behind it, so why is HTTP being blasted?
You might as well call it Exploit Pro Max...
I think we've all seen where an LLM names a session, issue, or PR based on its early understanding and fails update it later.
AI isn't being asked to review the title, and it doesn't have experience or think, so of course it's not going to flag it.
No human reviews anything anymore so of course they aren't going to catch it.
Edit: AI for as good as it is at "writing" is terrible at writing and particularly at naming things. I had an AI name a client OpenAICompatibleEmbeddingClient. "Brooooo.. I think "compatible" is implied, who names stuff like that?!"
Feels like the goal was sensationalism.
Just call it CVE-2026-111497. More accurate and much easier to remember and to talk about with your boss.
Or EUVD-2026-108169.
Actually it does, if the server is willing to decompress POST bodies.
Brotli bodies, so Brotli2Shell
It looks like marketing HTTP2Shell seems to appeal to everyone that uses HTTP, while Brotli2Shell might not wake up people that know what it is.
Now, whether it's warranted or not, that's another matter, but they know how to appeal for maximum awareness/interaction.
The author clearly used super intelligence.
I immediately thought HTTP/2 was having a bad day somehow. Might I suggest Software2Shell to be less confusing.
Specifically Brotli2Shell
google2shell would have been nice :D
Alt-tabbing from browser to terminal and back has always felt exhausting. Nice.
With ad served malware, this wouldn't be the first time though :)
All these names are silly. It's just marketing.
There are levels of silliness, and this has achieved a new one.
Much prefer the Broccoli2Shell name & site (for the same vuln): https://broccoli2shell.com/. For starters, I'm pretty sure you can hit this without HTTP if you do any manual Brotli decoding, HTTP just happens to be the most common case that comes up.
Why do critical vulnerabilities get a cutesy name and website nowadays? Who’s paying for these sites?
Are you asking why do we name things? In a programming forum?
> reported it as unauthenticated remote code execution that works in both directions
> but it wasn't recognized as a security issue at the time and no release shipped with the fix until now
I'm always getting my dangerous RCE bugs mixed up with my harmless RCE bugs.
I'm curious how this was reported and fixed back in Feb but nobody recognized the importance. A case of AI automation and nobody at the wheel?
Brotli2Shell seems to be a more appropriate name for this.
Not when you're prioritising sensationalism over accuracy. Everyone's heard of HTTP, but hardly anyone's heard of brotli.
call it GoogleBrotli2Shell then.
To me, the surprising part is how many of these implement brotli compression of request (not just response) bodies, given how obscure a feature that is.
In any case all this sensationalism is starting to get annoying. Patch your stuff and move on.
Edit: actually, between all the sensationalism, its hard to tell from the article how many of these actually support brotli compressed request bodies in common configs vs some other use of brotli (which is bad but nowhere near as critical)
I think the issue is that the servers "silently" support brotli. Then nobody uses it, but the attacker can still abuse the feature.
So as a hotfix I block this on my reverse proxy.
But is that actually true?
It seems like node supports it with body parser enabled.
But im having trouble finding any docs suggesting anything else supports it on inbound requests.
Obviously there are other scenarios where it is exploitable, but it is a much less severe vuln if the attacker must first force the program to make an outbound request to a server of its chosing rather than just connect to a vulnerable server.
The idea is to identify brotli encoding in the content-encoding header and block that on nginx. This shouldn't reach vulnerable brotli parser in node.
I hope my assets are not connecting to malicious servers, otherwise yes, I have a problem.
Asp.Net adds Brotli support if you "UseRequestDecompression" or "UseResponseCompression".
DotNet, however, has it's own C# Brotli implementation.
I wonder if we'll see more of this thanks to AI. Extremely complex features that nobody uses and that were easy to add, which incidentally tremendously expand the attack surface. One would hope that maintainer would be wise to avoid that, but for how long?
I think so, especially vibe coded software. I think hackers have the advantage now and defenders need to limit attack surface more rigorously.
Unfortunately that means understanding and limiting dependencies of developers... Yea, good luck.
My popcorn is ready. My servers are not.
Cloudflare has this amazing feature where it accepts compressed payloads and forwards them to unsuspecting backends. Can be weaponized in many ways.
A few month ago, everyone was advising to use cool-downs in your package upgrades. Now we're constantly overriding those because security incidents will be exploited immediately. Guess that's just how it's gonna be from now on.
It used to be a flex to say you deploy daily but it’s going to become a necessity just to keep up with the vulns
CI/CD — Continuous Incidents, Continuous Defense.
Article claims OpenAI found this in February, but I don’t think that’s correct. I think Google made a change to the Brotli source code in February that fixes the issue, but hadn’t realized it was a security issue and therefore hadn’t cut a release.
Yeah that doesn't make a lot of sense that they reported it but also nobody knew it was a security issue..
Maybe an LLM just happened to do one of those "while I'm here" fixes in Feb and because nobody reviews anything anymore..
"Fixed reading too many bytes."
LGTM!
Edit: I was close! "Detect and report invalid stream earlier." Could have been a person.
Well at least we know about it while or before the giant patch effort for everything else goes on. Just put it onto the pile.
Who turned on hard mode this week?
Nah man. We were on casual mode all along. This is the new normal.
Let’s say I have a node.js app behind an Amazon elastic load balancer which terminates TLS. My understanding is that in nearly all configurations, my app would be unaffected because the ELB would handle decompression. Is that generally correct?
I don’t know about that specific software, but I would be surprised if it decoded message bodies, which is entirely unrelated to TLS termination and shouldd be entirely unrelated to routing. So I would expect you would still be vulnerable.
Yep, I don't think too many people are exposing their python/node servers directly, it should be always behind a proxy or a fancy cloud LB.
Apache and nginx don't handle brotli out of the box.
I had to check all of our Python deployments and installation BOMs and was quite relieved - turns out that none actually installed brotli libraries. HTTP libraries like httpx, requests and frameworks consider them optional and autodetect their presence.
An example of the cost of added complexity I suppose. Brotli is much more complicated than the gzip/deflate it replaces. ... and of course we still have to expose gzip to attack as well.
Every time you add something you lose something...
I've just been pointing a Claude Code at these things as they pop up. A great time to not have to worry about chasing all these things manually.
wtf Python is also effected. https://pypi.org/project/brotli/ not sure about shell
Side point from going to pypi:
> Security notice: Upgrade certifi, pip, and uv in response to CA certificate incident
Blimey PIP is affected? That's bad. I feel like that should have been bigger news?
Because pip uses it's own certificate bundle, so even if you fix the system itself pip still needs its own upgrade to handle the eMudhra episode.
And strictly speaking you cannot trust the pip upgrade of pip with the vulnerable cert in the trust store due to certifi, which you also cannot upgrade safely with pip for the same reason. All of these are at one out of two compromises from being MITM'd during the upgrade.
So strictly we have to burn everything python to the ground and reinstall entirely.
Heh. This is part of why I always preferred getting my python runtime and all these core libs from my linux distro packager. But python die-hards look at me like I'm insane when I suggest this.
Even in today's LLM-addled world, I still think it's a better tradeoff for smaller system/site operators. Get the distro-level testing and regular auto-patching, rather than telling yourself it is a good idea to "freeze" your python world at the pace of your slowest app/system integrator.
It's the same reason I think docker is an anti-pattern for these small groups. You really cannot safely use these strategies to bake an image and divorce yourself from the maintained distro layer unless you are going to adopt CI/CD that does the same thing. Keep tracking all these patches to the entire runtime with all the dependencies, rebuild, re-test, and redeploy.
It's total cargo culting to adopt the freeze/image techniques and just abandon apps in place as if that is robust!
But, it's also more than just automatically updating everything. There is an editorial aspect to distro management to help secure the supply chain to your app. Most smaller orgs cannot feasibly do this as well as the regular linux distro maintainers. Some large orgs can probably do it better, but that doesn't matter if you aren't doing it large org style with dedicated security folks tracking all this against well-developed policies and procedures...
Or - you can just upgrade pip with a SHA hash pin, or set the magic env vars to use the system bundle.
You do need to be careful about creating new venvs with the old version of pip bundled in ensurepip in older Python releases. `uv` kindly patches that to newest pip, but not every Python binary does.
Or alternatively, on Debian with fixed system cert chain: https://discuss.python.org/t/avoiding-compromised-certificat...Python certificate handling is a hot mess, even by Python standards.
Pretty sure that's the eMudhra "private key is 4" issue.
How may web servers only serve brotli?
How many clients only accept brotli?
Well it's negotiated, right?
So say you run a SaaS that allows for something like letting people set their Avatar from an image URL which gets downloaded.
As an attacker I setup a new server that serves only brotli and then submit that URL to the SaaS.
Just send a POST with Content-Encoding:br to any URL and hope it gets decided before the URL route is checked.
> So say you run a SaaS that allows for something like letting people set their Avatar from an image URL which gets downloaded.
That's a clever example, but why would you do that? The avatar should be a static resource. As such, I'd only provide an upload file dialog.
Gravitar was super popular for a year or two.
Quit being overly pedantic about how contrived the example is and struggle with the underlying fact that the compression type is negotiated between the client and server. If the compression happens on the HTTP request (as some of the comments here suggest), then the server must use that compression method or it fails to read the body of the request.
Lots of apps will, for example, bring in your Google avatar (i.e. they fetch its URL for ingestion) as a default avatar if you use them for SSO. Slack included, IIRC.
Honestly it's time to get rid of C as a language used where security is a factor. Enough already. Why the hell is a lib like brotli that processes public general internet traffic, and maintained by google, not in rust?
I can understand why people used a C library for these apps, as when HTTP2 was specified, the C lib was the only game in down.
However, why not switch to the Rust lib when it was available? I suppose people at Dropbox are now pretty happy they did in 2016: https://dropbox.tech/infrastructure/lossless-compression-wit...
It's slower, though. Too high a price to pay for safety it seems.
Security is a factor everywhere. And C is everywhere. There's not going to be a "rewrite everything in Rust" project until after there is some legit penalty for selling / offering software in C that could be more secure in another language.
I think Mozilla got it right when they abandoned the effort to rewrite the entire browser in Rust and instead focused on analysis of which modules needed priority rewrites. If a module has historically had zero memory-related bugs, then there are other higher priority things to spend time on.
That said, Google Project Zero should have had Brotli on their radar. Hopefully they still have people and resources.
Because these applications were created long before Rust existed.
There is a cost to conversion.
Rust-Brotli (by Dropbox) is actually a drop-in replacement for the Brotli C lib: https://github.com/dropbox/rust-brotli#:~:text=rust-brotli%2...
Dropbox switched to it a decade ago, so it's probably quite product-ready.
Brotil was actually created post rust, but before rust took off.
It's not even an application it's a compression library
git for windows is vulnerable even in its latest release (2.56.0.windows.3), careful pulling repos over https from random ass servers (prolly uncommon but hey). they already merged the fix on the repo, it's just in a snapshot build only; should prolly arrive on Monday (2.56.0.windows.4).
also affects codex (gui), which ships its own git, could be more of a problem there, when the LLMs get quirky at night. still unlikely to be a problem though.
speaking of, update codex (and wipe its cache at ~/.cache/codex-runtimes), ships its own node. update python.
stuff like wireshark and imhex remain vulnerable, even on their latest repo commits, which may or may not be spicy.
also Microsoft 365 Copilot still has its own unpatched brotlidec.dll from Oct 7, if anyone uses that. no fix available for that one afaik, not sure about how it can be made to access the vulnerable decoder.
why is this downvoted? are any of those not accurate?
Because a local software vuln is much less scary, less exploitable, and more common than an exploit like PHP2Shell, which requires basically no effort to exploit a large chunk of publicly exposed webservers. Basically, local vulns have been a thing forever and GP is off topic.
I really don't understand how so many different things are affected by the same brotli bug/exploit.
I even read the patch for the C code (patch which, btw, seems a bit hacky to me with isSafe? methods added and whatnots but YMMV: now I understand things had to be patched in a hurry but this patch really looks like "lipstick on a buffer overflowing C pig" to me).
What's the link between Python / Node / C++ / libbrotli / Envoy / Bun: are those some all using the same C brotli encoder/decoder?
Can someone who knows more explain why all those are vulnerable?
> As for unaffected software, Apache’s standard mod_brotli does not decode Brotli
Because this one only encodes and doesn't decode and the bug is in the decoder? (for I take it mod_brotli, with such a name, still does something brotli related and isn't, say, a TOTP token generator nor a SAT solver?).
I mean overall, sheesh: could we get better explanation / better writeups when such big fuckups happen?
Yes. All these language runtimes compile against common libraries. Compression libs are one of the most common common libraries that platforms like node, ruby, dotnet, python, and etc need to compile.
Google provides the Brotli implementation. Package maintainers slurp that up and create ie libbrotli-dev. When you compile node it uses that for brotli support.
Brotli was a library released by the Google team that developed it. There weren't lots of independent implementations; they were all from the same source code.
Or you can roll your own and not rely on the literal disaster that Google developed. Brotli is trivial to implement.
As I've mentioned numerous times at this point, relying on third party code is a massive liability (supply chain attacks/sloppy code). My hand rolled brotli implementation is not only faster than libbrotli but also immune to this vulnerability (yes, I checked). Roll your own code people.
Would you also apply this to cryptography or, well, really any security library? Because I hope the answer will be no...
Why are we still so reliant on HTTP? HTTP came out in 1990 (36 years ago) and the most common version of HTTP still in use, 1.1 (RFC2616), came out in 1999 (27 years ago)?
Yes, I know HTTP 2 and 3 and SSE are attempts to make it faster and more parallel, but its still slow. Its slow because its unidirectional with a full round trip and by default its a new socket per each message. Yes, I know the real reason HTTP became and remains popular is because its sessionless... in theory. Still, there are better options available. You can do so much more with a full duplex channel.
We don't need new versions of http or SMTP because they're already simple and great and they're dumb protocols so there's not much you could ever add
If you want something else you can take your pick based on need from dozens of other options
HTTP 1 is full of parser differential vulnerabilities, which 2 fixes. That is probably its biggest flaw.
SMTP is too chatty - it uses too many round trips for efficient transfer across the internet. It can take several seconds just to get to the point of sending the email data. It also duplicates information between the SMTP envelope and the email body which creates vulnerabilities when they don't match.
For APIs that don't have to work in a browser I suggest JSON lines over TCP, or over TLS.
Most header parsing bugs that affect HTTP1.1 can also affect 2+3. Not in the same style but there are quite a few overlapping ones.
"by default its a new socket for each message" is misinformation - keepalive is the default since HTTP 1.1 (1999).
Agreed. The issue with all these archaic standards is that they were rolled decades ago back when certain practices were either outright not being employed OR that hardware wasn't fast enough. Ironically, HTTP itself is a literal nightmare to parse (performance wise) on top of having like... ten? distinct vulnerability classes. The world desperately needs new standards.
Just open a TCP socket and send JSON messages. Or if you want a bit more efficiency, fixed-format binary messages, but remember to include a version number.
For custom server implementations, agreed. But the web still depends on HTTP and we sadly need the major infra providers to agree on a new standard otherwise HTTP isn't going anywhere