Safety7d ago
Safety
Risk Reports Must Address Deployment Misalignment
Pre-deployment alignment assessments fail to capture motivations that emerge only after an AI is active. The AI Alignment Forum argues that dangerous goals can develop during deployment even if initial training was benign. Evaluators must integrate this specific risk into planning. Ignoring this gap leaves companies unable to disprove the possibility of consistent adversarial misalignment.