Safety
104d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments often fail to capture motivations that emerge only after a model is released. AI Alignment Forum argues that benign models can develop dangerous goals during active deployment. This gap makes current risk analysis insufficient.
The Signal
Evaluators must incorporate deployment-time spread to prevent consistent adversarial misalignment in frontier models.