OpenAI temporarily slowed its model scaling pace after an incident involving Hugging Face and preliminary evidence that its upcoming model Astra may meet a critical cybersecurity capability threshold. The company paused reinforcement learning training for two weeks to harden research environments, expand monitoring, and conduct red-teaming. OpenAI now requires stricter security standards, including workload and network isolation, for Astra and cyber-related workloads, with several research runs remaining on hold until fully migrated.
No score is assigned. Sources and their independence are shown in the citation chain below.