9.2
CVSS 4.0 score Google assigned to CVE-2026-111497
11
public GitHub repositories for the CVE a day after the fix, per Sysdig
534 bytes
of input that hung pyarrow in JFrog's test
236 days
between the upstream fix commit and the first release carrying it
Nobody installed Brotli on purpose
Almost no one reading this chose to install Brotli. It came with something else. Node.js bundles it. .NET ships it. Envoy and Istio compile it in. PHP's Windows builds carry it. The Python packages that make requests and aiohttp understand compressed responses pull it in quietly. When Google's Brotli decoder turned out to have an out of bounds write, CVE-2026-111497, the fix had to travel through every one of those products separately, on its own schedule, under its own version number.
That is the argument of this post. The patching for CVE-2026-111497 is not hard. Finding everything that needs it is. A vulnerability scanner that reads product versions will tell you to update Node.js and .NET, and it will be right. It may not tell you that a self contained .NET app you deployed in the spring still carries the old runtime inside it, that your PHP on Windows reports the same version before and after its fix, or that a data pipeline is reading Parquet files through a copy of Brotli 1.0.9 that Apache Arrow bundles.
We are not saying the bug is being exploited. CISA's own assessment says it is not, as of 23 Oct, and we will be exact about that below. We are saying this bug is a clean test of whether your inventory knows about libraries, or only about products. Most inventories know about products. Attackers aim at libraries, because one library bug is a key to many doors.
What CVE-2026-111497 is and who fixed it
The flaw lives in the Brotli decoder, c/dec/decode.c, in Google's C library, versions 1.0.3 through 1.2.0. The NVD record describes a buffer overflow that lets "a remote unauthenticated attacker" corrupt decoder memory, disclose memory contents, cause a denial of service, "or potentially execute arbitrary code using a crafted Brotli-compressed stream." Google, as the numbering authority, scored it 9.2 under CVSS 4.0: network reachable, low complexity, no privileges, no user interaction, with one qualifier, attack requirements present. The requirement is simple to state. The target has to decompress Brotli data the attacker supplied.
The Node.js advisory gives the clearest plain account of the mechanism. An invalid block switch in the compressed stream was reported as a temporary need for more input instead of a permanent format error, so the streaming decoder kept appending input bytes to an eight-byte internal buffer. Past the end of that buffer sits other decoder state. Node.js names the consequences in order: a crash, disclosure of process memory, or remote code execution.
Google released fixed Brotli versions 1.0.10, 1.1.1 and 1.2.1 on 21 Oct 2026. Node.js shipped 22.23.4, 24.22.1 and 26.12.1 the same day, crediting OpenAI's Codex Security team with the report. Over the next five days the fix spread outward: Bun 1.4.4 on 22 Oct, Envoy and the Python package advisories on 23 Oct, Microsoft's out of band .NET updates on 25 Oct, and rebuilt PHP Windows packages on 26 Oct.
-
Component //
Google Brotli decoder, c/dec/decode.c -
Affected //
Brotli 1.0.3 through 1.2.0, and every product that bundles them -
Fixed upstream //
1.0.10, 1.1.1, 1.2.1, released 21 Oct 2026 -
CVSS //
9.2 (4.0, Google); 8.1 (3.1, Red Hat) -
Trigger //
the target decompresses attacker supplied Brotli data -
Exploited //
No, per CISA's assessment of 23 Oct
Two scores circulate and both are honest. Red Hat rates the flaw Important at 8.1 under CVSS 3.1, setting attack complexity to high, and notes that hardening such as ASLR and NX makes reliable code execution harder while crashes stay easy. Use the higher number to set urgency on anything internet facing that decompresses untrusted input, and the lower one as a reminder that a crash is the likely outcome of a failed attempt.
One date in the history is worth knowing. The fix commit in Brotli's repository, 0ed4268, "Detect and report invalid stream earlier," is dated 27 Feb 2026. The last release before it, 1.2.0, came out in October 2025, and no tagged release carried the fix until 21 Oct 2026, 236 days after that commit. Anyone who builds from tagged releases, which is nearly everyone, had no way to pick it up. We made a related point on 25 Oct about a Collabora fix that shipped before its CVE existed. Here the order is reversed: the CVE is loud, and the fixed copies are the hard part to see.
Is CVE-2026-111497 being exploited?
No exploitation has been reported. CISA's coordinator assessment on the NVD record, dated 23 Oct, reads exploitation "none," automatable "no," technical impact "total." The flaw is not in the Known Exploited Vulnerabilities catalog as of the feed released late on 26 Oct. We checked.
What has happened is faster than usual. Imperva's red team demonstrated code execution against the official Node.js 24.22.0 image on Linux arm64 on 22 Oct and, according to Sysdig, has not released its exploit. That is one architecture and one release line, not a universal weapon. Sysdig also counted public activity: the first GitHub repository for the CVE appeared about five hours after the fixes shipped, and a search at 20:28 UTC on 22 Oct returned 11. Red Hat's position is that there is "no public, reliable remote code execution proof-of-concept." Both can be true. Crash inputs are easy and public. Reliable code execution is hard and private, for now.
For a defender that maps to a clear stance. Treat any exposed service that decompresses Brotli as a crash target today, which for many services is a denial of service problem in its own right, and as a code execution target soon. The interval between a public crash and a working exploit for a memory corruption bug is set by how much an attacker wants it, and a library this widely embedded is worth wanting.
Where Brotli hides, and how each copy gets fixed
This is the part a product scanner tends to flatten. Each row below is a different fix path, and several of them change nothing a version check would notice.
Node.js. Node.js bundles its own Brotli, and the vulnerable path is node:zlib. It is reachable wherever an application decompresses Brotli it did not create: a fetch() call that accepts a compressed response, or server side middleware that decodes a Content-Encoding: br request body. JFrog's write up says Imperva used both. A plain Node.js HTTP server does not decode request bodies by itself, so exposure depends on the framework. Lines 25.x and 20.x and older are end of life and get no fix. Distribution packages complicate it further. Red Hat says Node.js 24 and 26 on RHEL 8 and 9 use the bundled Brotli, while Node.js 22 there uses the system library. The same version string can mean a different copy depending on who built it.
Bun. Bun 1.4.4, released 22 Oct, lists Brotli among its security fixes. Anything on 1.4.3 or earlier should move.
.NET. Microsoft fixed 8.0.33, 9.0.22, 10.0.14 and 11 RC3 on 25 Oct. The exposure is through ASP.NET Core request decompression, HttpClient automatic response decompression, or the System.IO.Compression Brotli APIs. The first two are off by default, which narrows the field. The trap is deployment. Microsoft says self contained applications must be "rebuilt and republished with the updated runtime," because installing a new runtime on the host does not touch the copy inside the app. The same goes for container images. A patched server can run an unpatched app.
Envoy and Istio. Envoy fixed 1.40.0, 1.39.3, 1.38.6, 1.37.8 and 1.36.12, and its advisory limits impact to "deployments that have Brotli decompressor configured." Istio's fixed releases are 1.31.2, 1.30.6 and 1.29.9, aimed at users who enable Brotli decompression, for example through an EnvoyFilter. If your mesh never decompresses Brotli, the bug is present but unreachable. If one team added a decompressor filter two years ago, the edge proxy is the most exposed copy you own.
PHP on Windows. On 26 Oct PHP published rebuilt Windows packages, 8.5.12-1, 8.4.27-1, 8.3.36-1 and 8.2.35-1, carrying Brotli 1.2.1. The PHP announcement is explicit that "the PHP runtime versions remain unchanged." php -v reads the same before and after. Only the package suffix and the file hash tell them apart. The Hacker News reports the path is the cURL extension with automatic decompression enabled, which is off by default. This is the purest example of the problem: a fix that a version based scanner cannot see by construction.
Python. The PyPI Brotli package compiles the same C decoder; 1.2.1 is fixed, and brotlicffi is fixed in 1.2.1.0. The Python advisory adds the line that matters most for exposure: with Brotli installed, common HTTP clients decode Brotli responses automatically, naming urllib3 and requests, httpx with the brotli extra, and aiohttp. Installing the package is enough to turn on the code path. Every outbound request your Python service makes to a server it does not control is a way in.
The long tail. JFrog's research on 25 Oct went past the runtimes. It crashed a ClickHouse server over HTTP with one crafted 23 MB request body. It reached the decoder in FFmpeg through libjxl, the JPEG XL library, and showed instruction pointer control, while stating that full code execution "would require further exploitation." It hung pyarrow in an endless loop with a 534 byte input, because Apache Arrow bundles Brotli 1.0.9. JFrog's warning is the thesis in one line: Brotli arrives through "transitive dependencies, bundled into native binaries," which version only scanners may miss.
Here is how input reaches the decoder in all of these, at the level of what each step achieves.
Figure 01
From compressed bytes to a corrupted process
- 01
Deliver //
Crafted Brotli data arrives as a request body or a server response.
- 02
Auto decode //
A framework or client decompresses it without asking.
- 03
Overrun //
A bundled decoder writes past its buffer.
- 04
Outcome //
A crash today, memory disclosure or code execution where it works.
Result //
The product you patched is not always the copy that decoded the bytes.
Notice that step 01 runs in both directions. Server side, the attacker sends a compressed body to your API. Client side, your service fetches a URL and the attacker controls the response, which covers webhooks, link previews, feed readers, scrapers and any integration that calls a third party.
How to find Brotli copies and detect Node2Shell attempts
Patch the products first. Then do the work this bug is really about, which is finding the library.
Start with what your build systems already know. Lockfiles and SBOMs will show the PyPI Brotli and brotlicffi packages, Node.js and .NET versions in container images, and Envoy and Istio versions in your mesh. That covers the visible rows.
Then go below the package layer. JFrog recommends "examining source code and binaries to find relevant decoder functions," because a vendored copy has no package entry. In practice that means searching shared libraries and executables on your hosts and inside your images for Brotli's decoder symbols, such as BrotliDecoderDecompress, then deciding for each hit whether it is linked to a fixed system library or carries its own copy. JFrog names an open source scanner that does this for ELF binaries, and Trail of Bits has published a checker that fingerprints vulnerable source and binaries. For PHP on Windows, compare the package file name and hash against the PHP downloads page, not the version string. For self contained .NET apps, the question is when the app was last published, not what the host has installed.
Then find the decoders that are switched on. Search configuration for Brotli decompressor filters in Envoy and EnvoyFilter resources in Istio, ASP.NET Core request decompression middleware, HttpClient handlers with automatic decompression that includes Brotli, Node.js middleware that decodes br bodies, and PHP cURL options that enable automatic decoding. A copy that is present but never fed untrusted input can wait for the normal patch cycle. A copy behind an enabled decoder on an internet facing path cannot.
For the edge, Cloudflare added two managed WAF rules on 25 Oct, one for requests that declare Content-Encoding: br and one for the Next.js image optimizer path. Both default to Log, not Block. That makes them a useful source of telemetry: very few legitimate clients send Brotli encoded request bodies, so a burst of them is worth a look. It also means they protect nothing until someone changes the action.
On hosts, the signal Sysdig points to is the crash. Failed attempts crash the Node.js process, and a successful one tends to be followed by something a web server process should not do. The rule below watches for a Node.js process starting a shell, a starting point that fits this bug and most other server side Node.js compromises.
Figure 02
Node.js server process starts a shell
Sigma
title: Node.js process spawns a Unix shell
status: experimental
description: A Node.js process starts a shell. Web facing Node.js services rarely do this, and it is the common next step after code execution through flaws such as CVE-2026-111497.
references:
- https://nodejs.org/en/blog/vulnerability/october-2026-security-releases
- https://nvd.nist.gov/vuln/detail/CVE-2026-111497
logsource:
product: linux
category: process_creation
detection:
selection:
ParentImage|endswith: '/node'
Image|endswith:
- '/sh'
- '/bash'
- '/dash'
- '/zsh'
condition: selection
falsepositives:
- Build tools, test runners and npm lifecycle scripts on developer and CI hosts
- Applications that shell out by design; baseline them first
level: high
tags:
- attack.execution
- attack.t1059.004
- cve.2026-111497 How we test for bundled library exposure
On a sanctioned engagement, a library flaw like this changes how we read a target, not just which version numbers we check. The question we ask is where the target decompresses, parses or renders data it did not produce, and which code does that work. Compression, image decoding, document parsing and archive extraction are the usual places, and they are almost always handled by a native library the application team did not choose.
From the outside, that starts with mapping inputs. Which endpoints accept compressed request bodies, which accept file uploads that get converted or thumbnailed, which features fetch a URL the user supplies. Our web application recon guide covers that mapping step, including checking every version string against advisories, and the warning that comes with it: a version string is a claim, not a fact.
From the inside, on an assumed breach or internal test, we compare what the asset inventory says with what is on disk. On our engagements the gap is reliably at the library layer: container images built once and never rebuilt, applications that ship their own runtime, appliances whose firmware bundles open source components the vendor does not list. None of that shows up in a scan that reads product banners. All of it shows up when you look for the code itself. We made the same point about PHP builds whose version strings could not be trusted, and the tools we publish follow that rule: check what is installed and how it is configured, not what it says about itself.
We do not need to fire a memory corruption exploit to make the finding. The finding is that an internet facing service decodes untrusted Brotli through a vulnerable copy that the patch process did not know about. Proving reachability and the vulnerable build is enough for a client to act, and it avoids crashing production to make a point.
What is still coming for shared decoders
Three things to expect.
More copies will surface. JFrog's list is a sample, not a census. Any software that accepts Content-Encoding: br, decodes WOFF2 web fonts, reads JPEG XL images, or handles Parquet and Arrow data may carry Brotli, and the fix will reach it at the speed of each maintainer. Appliances and embedded products are the slowest of all, and the hardest to inspect.
Reliable exploits will follow the crashes. Today there is one private demonstration of code execution on one platform and many public crash inputs. That ratio changes. When it does, the bug will hit the services that were left alone because the CVSS score was argued down or the advisory said "potentially."
And AI assisted bug hunting will keep landing on shared libraries. OpenAI's team found this one, and a decoder used by half the internet is exactly the kind of target automated review is good at. Each such find produces the same pattern we see here: one upstream fix, then weeks of downstream rebuilds. Defenders who can answer "where does this library live in our estate" within a day will handle that pattern well. Defenders who can only answer "which products do we run" will repeat this exercise every month.
What to do today
-
Patch the runtimes and proxies you know about. Node.js 22.23.4, 24.22.1 or 26.12.1; Bun 1.4.4; .NET 8.0.33, 9.0.22, 10.0.14 or 11 RC3; Envoy and Istio fixed releases; PyPI Brotli 1.2.1 and brotlicffi 1.2.1.0; PHP Windows builds with the -1 suffix. Move anything on Node.js 20 or 25 to a supported line.
-
Rebuild what carries its own copy. Republish self contained .NET apps and rebuild container images after the host and base image are fixed. A patched host does not patch a bundled runtime.
-
Find the decoders that are switched on. List every Brotli decompressor filter, request decompression middleware and HTTP client with automatic Brotli decoding. Turn off the ones nobody needs, starting with internet facing paths.
-
Search binaries, not just manifests. Look for Brotli decoder symbols in images and hosts, and treat a vendored copy you cannot update as a reason to stop that service decoding untrusted input.
-
Watch for crash loops and shells. Alert on repeated Node.js restarts and on Node.js processes starting shells; the Sigma rule above is a starting point. If you use Cloudflare, review the new Brotli rules, which only log by default.
SOURCES
- NVD // the CVE description, affected Brotli 1.0.3 to 1.2.0, Google's CVSS 4.0 score of 9.2 and CISA's no exploitation assessment
- Node.js October 2026 security releases // the bundled decoder flaw, affected lines and the remote code execution impact
- JFrog Security Research // the flaw reached through ClickHouse, FFmpeg, .NET and pyarrow, and why version only scanners miss bundled copies
- .NET Blog // fixed .NET versions and that self contained apps and containers must be rebuilt
- PHP // Windows rebuilds with Brotli 1.2.1 under unchanged PHP version numbers