Node2Shell: Brotli RCE triggers on any fetch() — patch Node now
A malformed Brotli response can leak memory or run code in Node 22/24/26 — PoC out, no workaround. The same patch also kills TLS/DNS spoofing bugs.
Copy markdown
The bug: an out-of-bounds write in Node's Brotli decoder
CVE-2026-111497 lets malformed Brotli input corrupt memory when Node decompresses it via node:zlib — the outcome is a crash, process-memory disclosure, or full remote code execution. It affects every supported line (22.x, 24.x, 26.x) and all end-of-life builds.
Why fetch() makes this everyone's problem
The decoder fires automatically when fetch() (built on undici) receives a Brotli-compressed response — so any app that calls an untrusted or attacker-controllable endpoint is exposed, no explicit zlib call needed. axios and other undici clients inherit it. There's no workaround: upgrade to 22.23.4, 24.22.1, or 26.12.1.
A PoC is out — patch today, but don't panic-deploy
A public proof-of-concept landed Oct 22. It only reached code execution against a deliberately weakened build; on a standard build the author got a crash and a limited memory read, with no in-the-wild exploitation reported. Roll the upgrade on your normal cycle — just don't let it slip.
The same release quietly closes TLS/DNS spoofing holes
While you bump versions, the October drop fixes three High-severity cert-validation bypasses: DNS names truncated at a NUL byte, fullwidth-digit IPs, and leading-zero IPv4 literals can each make a certificate match the wrong host. If your service makes HTTPS calls, these matter as much as Brotli.
Build-snapshot users: your randomUUID is frozen
CVE-2026-106074 bakes the caches behind crypto.randomUUID(), randomUUIDv7(), and randomInt() into --build-snapshot and single-executable (SEA) binaries, so every launched process replays the same "random" values. If you ship snapshots or SEAs, rebuild after upgrading.
Agent Skills joins a vendor-neutral foundation
Anthropic donated Agent Skills to the new Agentic AI Foundation under the Linux Foundation, alongside OpenAI's Open Responses spec, Hugging Face's OpenEnv, and Alibaba's OpenSandbox. With 70+ platforms already integrated, the skill format you write for Claude is becoming portable across clients.