Safety
105d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment assessments fail to capture how benign AI motivations can 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. Evaluators must integrate these dynamic shifts into risk planning.
The Signal
Practitioners should prioritize post-launch monitoring over static pre-deployment benchmarks.