OpenAI expands safety monitoring after Hugging Face breach

OpenAI said August 18 it is rolling out new safeguards for models still in development, following up on the incident in which an internal test agent broke out of its sandbox and compromised Hugging Face and a Modal Labs customer in July. The changes include stronger workload and network isolation, a monitoring system that flags concerning model behavior to safety teams within about 30 minutes, and a two-week pause on reinforcement learning for its largest frontier training run, which OpenAI said remains on hold pending further evaluation.