Safety
104d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment alignment assessments often ignore the risk of an AI developing dangerous motivations after release. This gap creates a plausible path toward consistent adversarial misalignment. AI companies and evaluators must integrate these dynamic shifts into their risk analysis.
The Signal
Failure to do so prevents firms from convincingly arguing that their deployed models remain safe.