Pacing model development in an era of cyber-critical capabilities

OpenAI has announced a series of stronger monitoring, alignment, and security safeguards for its frontier AI models in response to growing risks. The changes follow the OpenAI-Hugging Face incident and preliminary evidence that an upcoming model, Astra, may meet the Critical cybersecurity capability threshold under OpenAI's Preparedness Framework. To ensure standards stay ahead of growing risks, OpenAI temporarily slowed the pace of scaling, including a two-week pause in reinforcement learning training on its latest models intended for deployment, and paused frontier model inference in research clusters immediately following the Hugging Face incident. OpenAI has defined and started implementing stricter security requirements, which include workload isolation using sandboxes, network isolation to separate higher-risk workloads from the internet, and continuous security testing to remove vulnerable shared services. Additionally, the company introduced a multistage monitoring setup for models of Sol capability or higher, utilizing activation classifiers and automated investigators to scan for unauthorized access, data theft, and attempts to defeat safeguards. This monitoring incurs an estimated compute overhead of roughly 20 percent. While some Astra training and evaluations meet these new requirements, a significant number of workloads remain paused until fully migrated. OpenAI also emphasized that alignment research remains central to its mission, noting that advanced capabilities like cyberattacks require stronger evidence of aligned behavior throughout all stages of training.