Frontier AI Risk Assessment Framework

This technical report outlines NVIDIA's risk assessment framework for frontier AI, which is applicable from August 2025. It describes the structured approach to identifying, mitigating, and addressing potential harms associated with advanced AI models. The report defines 'frontier models' as highly capable general-purpose AI models that can perform a variety of undefined tasks, posing unpredictable risks if not managed responsibly. The risk framework consists of two main components: the Preliminary Risk Assessment (PRA) and the Detailed Risk Assessment (DRA). The PRA categorizes AI models based on their intended use, capabilities, and level of autonomy, assigning a model risk score. The DRA evaluates the product's architecture and development processes, integrating risk assessment into the product lifecycle. The report emphasizes the need for continuous review and adaptation of the risk framework to reflect new findings and threats in the AI landscape, ensuring that risk assessments remain relevant and effective over time.