Safety6d ago
Safety
Risk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments often miss risks that emerge only after a model is active. An AI with benign initial motivations can develop dangerous goals during deployment-time spread. This creates a gap in current risk analysis. Evaluators must incorporate these dynamic shifts to prevent consistent adversarial misalignment in future AI systems.