Skip to content
SectorFrame

Independent intelligence on AI

OpenAI’s sandbox drill finds a critical hole in the web’s compression plumbing

Agents set to attack OpenAI’s own training enclosure found an unauthenticated code-execution bug in Brotli, the library inside Node.js, Envoy and Python. The fix had sat in the source code since February, unreleased and unlabelled.

Topics
Regions
AI capabilities
Keep SectorFrame in your search

Brotli is the kind of software nobody thinks about, which is precisely what makes a flaw in it dangerous. Google’s compression library sits inside Node.js, Envoy, Bun, Python packages and some builds of curl, quietly shrinking web traffic. On 23 October OpenAI said that agents it had set to attack its own research infrastructure, after a model escaped its training environment through a DNS error in September, had found a bug in Brotli’s decoder that allows unauthenticated remote code execution. It is now CVE-2026-111497, rated 9.2, or critical, under CVSS 4.0.

This article continues for subscribers. Subscribe to read the full analysis.

SectorFrame membership

Read the full story

Independent reporting for people putting AI to work. Sign in to check your access, or explore a subscription.