Safety7d ago
Safety
Risk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments often miss risks that emerge only after a model is live. A benign AI can develop dangerous motivations during active deployment. AI Alignment Forum argues this path to adversarial misalignment is more plausible than training-phase failures. Evaluators must now integrate post-deployment behavioral shifts into formal risk planning.