OpenAI's security review uncovered a flaw in Brotli, a compression library used in Node.js, Bun and other software, after a DNS filtering lapse gave an internal model limited access to a public chatbot.
The incident occurred Sept. 20, when a resolver gap let the model query a public chatbot. OpenAI said other requests went to an offline web cache, so the model did not reach other live sites. The company attributed the incident to inadequate DNS filtering.
OpenAI paused training and evaluation of its most capable models during security tests, then resumed frontier reinforcement-learning training and internal inference on Oct. 1. EconoTimes previously covered the restart after the DNS incident.
The security sweep identified CVE-2026-111497. GitHub's advisory gives it a 9.2 score under CVSS v4.0 and classifies it as Critical; the Node.js Project rates its advisory High. The Node.js notice credits OpenAI Codex Security with reporting the flaw.
A specially formed compressed input can make the decoder write beyond its allocated memory, creating a risk of code execution in software that processes it. Successful exploitation is not assured in every deployment. Named affected software includes Node.js, Bun and the Python packages Brotli and brotlicffi.
Brotli and Node.js issued fixes Oct. 21, while Bun, Envoy and Istio published related updates. A February change in Brotli's upstream code had addressed the issue, but it was not identified as a security fix then, so dependent projects needed separate updates.
OpenAI cited a July incident at Hugging Face as a precedent. The company said findings from its red-team work would inform safety cases for future training and evaluation.
