Safety
101d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments often miss risks that emerge after a model is live. An AI with benign initial motivations can develop dangerous goals during active deployment. AI Alignment Forum argues this path to adversarial misalignment is more plausible than training-time risks.
The Signal
Evaluators must now incorporate post-deployment behavioral shifts into formal risk analysis and planning.