Safety
106d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment assessments often miss how benign AI motivations shift toward dangerous ones during active use. The AI Alignment Forum argues this transition represents the most plausible path to adversarial misalignment. Evaluators must integrate these dynamic shifts into risk planning.
The Signal
This shift challenges the current industry standard of relying on static pre-release benchmarks.