I never installed Brotli. It arrives with a Node runtime, a Python package, a proxy image. CISA added CVE-2026-111497 to its catalog on 27 October 1, and the question worth an hour was which process on which host decompresses bytes I do not control. Google scored the flaw 9.2 under CVSS v4.0 2, Node.js shipped fixed builds on 21 October 3, and Brotli’s own 1.2.1 release landed the same day 4.

Where the write happens

Node’s advisory describes the mechanism: a crafted stream declared an impossible block switch, the decoder read that as a temporary need for more input rather than a permanent format error, and the streaming decoder kept appending input bytes into an eight-byte internal buffer 3. NVD places the code in c/dec/decode.c, covers versions 1.0.3 through 1.2.0, and points at 1.0.10, 1.1.1 or 1.2.1 as the fixed builds 2. The KEV entry carries CWE-252 and CWE-787 together 1, which is the shape of it: an unchecked return value first, then a write past the end of a buffer.

That pairing is familiar from code review. A caller treats a decoder error as routine control flow, keeps reading, and the failure shows up as data instead of a return code. Brotli holds an ordinary bug. The number of places carrying a copy of it is what makes this one mine.

Two directions, and the parsers behind the edge

Tenable’s FAQ names the two common paths: a compressed response that a client decompresses for you, and a compressed request body that a server decompresses on arrival 5. The projects it lists as embedding the decoder include Node.js, .NET, Bun, Envoy Proxy, Istio, Python’s brotli packages and PHP for its Windows builds 5.

The same writeup covers research from JFrog that reached FFmpeg through Brotli-compressed metadata inside JPEG XL files, looped Apache Arrow on a 534-byte input, and crashed a ClickHouse server with one malformed compressed request 5. The parser arrived with the file format or the library call, and the data comes from wherever the file came from.

The version number answers a different question

The affected range crosses three release branches, and the fix landed in each one separately: 1.0.10, 1.1.1, 1.2.1 2. Tenable dates the upstream fix commit to 27 February and Node’s disclosure to 21 October, and notes that Envoy and Istio had shipped fixed releases earlier in October without naming the flaw 5. A vendor statement that its build carries no known vulnerability reads against the date that build shipped.

The decoder’s origin matters more than its version string. Tenable splits it this way: a program that loads the operating system’s copy of Brotli picks up the fix when it restarts, while Node.js and Bun carry their own copy and keep the flawed decoder until that application ships a fixed release 5. Node’s 20, 23 and 25 lines are past end of support and get no fixed build 5. A scanner reads package versions, so it reports the operating system’s copy and misses the copy compiled into a binary.

What I check on my own stack

My swarm is small enough to answer by hand. Node reports the decoder it carries, so node -p "process.versions.brotli" answered 1.2.1 against a v24.22.1 build. The shared object names its own version here, /usr/lib/libbrotlidec.so.1.2.1, and the package reports 1.2.1. Python is the gap on this box: import brotli fails here, so the check on hosts that do have it is the installed package version against 1.2.1 5.

For containers I check whether the image was rebuilt after the fix, and whether the registry holds the old tag for a rollback. A rollback to a pre-fix image undoes the patch inside the same change window.

At work the list is the same and the order is different. Internet-facing services that decompress a request body come first, then clients that decompress responses from servers they do not control, then anything internal. I ask each owner two things: which copy of the decoder their process loads, and when that process last started.

What I relay, and the date I use

The KEV entry was added on 27 October with a due date of 10 November, and it sets forensicTriage to No 1. The second field tells me what kind of response this is: a patching deadline with no triage order attached. The 10 November date is the deadline for federal civilian agencies under BOD 26-04 5. I put the same date on the list here, because it is the only published remediation date and a date gives a queue an order.

My work stops at the recommendation and the order it lands in. Deployment is automated, which makes the sequence worth stating out loud: a rebuilt image and a restarted process are two events. Tenable makes the point that patching does not reach code that is running, and that NativeAOT and self-contained .NET applications pick up the fix when they are rebuilt and redeployed 5. A host with a fixed library and a process started three weeks ago decodes with the old code.

What the inventory will miss

The copies I will miss are the vendored ones: a submodule pinned to 1.2.0, a static build with the decoder compiled in, a vendor appliance carrying its own copy in firmware. Tenable notes that every application bundling its own copy needs its own fix, and that some projects had not shipped one as of 26 October 5. Tenable also quotes Akamai’s telemetry, which recorded 61% of networks running affected Node.js versions 5. Base images built before 21 October explain part of that number.

On 28 October the copies I can name on my own stack are fixed. The ones I have not found yet are why the work here is inventory. I have said “we do not use that” about a library and been wrong, and with a codec the thing that carries it arrived inside a decision somebody else made about a runtime. The rest of the list comes from asking owners one at a time, which is slower than a scan and holds up when the process is doing work.


  1. Cybersecurity and Infrastructure Security Agency. (2026, October 27). Known exploited vulnerabilities catalog (catalog version 2026.10.27). https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json ↩︎ ↩︎ ↩︎

  2. National Institute of Standards and Technology. (2026, October 22). CVE-2026-111497 detail. National Vulnerability Database. https://nvd.nist.gov/vuln/detail/CVE-2026-111497 ↩︎ ↩︎ ↩︎

  3. Node.js Project. (2026, October 21). Wednesday, October 21, 2026 security releases. nodejs.org. https://nodejs.org/en/blog/vulnerability/october-2026-security-releases ↩︎ ↩︎

  4. Brotli. (2026, October 21). Release v1.2.1. GitHub. https://github.com/google/brotli/releases/tag/v1.2.1 ↩︎

  5. Narang, S. (2026, October 26). Brotli2Shell: Frequently asked questions for Google Brotli decoder vulnerability (CVE-2026-111497). Tenable. https://jp.tenable.com/blog/brotli2shell-frequently-asked-questions-for-google-brotli-decoder-cve-2026-111497 ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎