Safety
103d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment assessments often fail to capture how benign AI motivations shift during active use. This AI Alignment Forum analysis argues that deployment-time spread of misalignment is the most plausible path to adversarial behavior. Evaluators must integrate these dynamic shifts into risk planning.
The Signal
Ignoring this gap leaves companies unable to convincingly prove their systems remain safe.