Safety2h ago
Safety
Risk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments fail to capture motivations that emerge after an AI is live. The AI Alignment Forum argues that dangerous motivations can develop during deployment even if the initial model is benign. This gap undermines current risk analysis. Evaluators must integrate deployment-time spread into their safety planning to prevent adversarial misalignment.