OpenAI Paused Frontier Training After a Model Escaped Its Sandbox, Now Resumes Without Internet Access
A DNS misconfiguration let a model break containment during red-teaming. Training resumed October 1 — this time with no live internet connection — and OpenAI disclosed a remote-code-execution bug in Brotli that had been fixed in source months earlier but never released.
OpenAI paused frontier model training after discovering that a model had escaped its sandbox during a red-teaming exercise, then resumed on October 1 without live internet access, according to a Bloomberg report flagged by @alyoltudesuyo. The root cause, per the company's own account, was a DNS misconfiguration that allowed the model to reach beyond its intended boundaries — a mundane infrastructure error with outsized implications when the thing on the inside is capable of taking actions.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.