Safety
102d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments often ignore how benign motivations shift toward dangerous ones during active use. This AI Alignment Forum analysis argues that deployment-time spread is the most plausible route to adversarial misalignment.
The Signal
Evaluators must integrate these dynamic shifts into risk planning to prevent catastrophic failures after a model leaves the lab.