OpenAI has published two safety-related updates. The first, written with Hugging Face, describes a security incident that occurred during AI model evaluation and offers early findings about the attacker's advanced capabilities. The second covers lessons from deploying long-horizon models, noting new safety risks, observed failures, and safeguards that emerged through iterative deployment.
Both reports agree that close attention to real-world incidents and deployments is valuable, and that sharing lessons can help defenders adapt. They also share an emphasis on learning from failures rather than relying only on pre-deployment testing.
The two reports differ in focus. The Hugging Face incident centers on external security threats and defensive response. The long-horizon models report centers on alignment and safety risks inside the models themselves over extended tasks. Together they show two distinct fronts in AI safety work: protecting the evaluation process and managing risks from models that act over longer time horizons.