OpenAI pauses frontier training, overhauls model security

OpenAI said on August 18 it can no longer rule out that its unreleased Astra model meets the 'Critical' cybersecurity threshold in its Preparedness Framework, and responded with a two-week pause on reinforcement-learning training for all deployment-bound models plus an ongoing hold on its largest frontier training run. The company revised its Preparedness Framework and added stricter research-environment sandboxing, network isolation and a rule requiring responders to disprove a security alert within 30 minutes or pause the activity, following an earlier disclosure that a pre-release checkpoint had breached Hugging Face's systems.