Over the past weeks, two incidents highlighted growing risks from capable AI systems: the Open AI-Hugging Face incident and evidence that our upcoming model Astra may meet the Critical cybersecurity capability threshold under our Preparedness Framework. These developments, alongside rapid internal research progress, added urgency to strengthening monitoring, alignment, and containment safeguards across all training stages.

As models grow more capable, so do associated risks. We temporarily slowed scaling, including a two-week pause in reinforcement learning (RL) training, to harden research environments, red-team systems, and expand monitoring coverage. Our largest frontier RL run remains on hold while we conduct smaller-scale training and evaluations to validate safeguards and establish alignment evidence.

We now require stronger evidence of aligned behavior throughout training, building on ongoing research and evaluations. The signals from upcoming models make clear the need for a broader approach beyond the current Preparedness Framework. Below, we detail changes to research processes and infrastructure, and the work still underway.

Our approach rests on three reinforcing safeguards: monitoring, alignment, and security. We apply these across research and deployment, adapting them to each model's capabilities, operating environment, and risk level. Frontier models gaining stronger cybersecurity capabilities require raising security standards for training and evaluation environments.