Safety
104d agoRisk Reports Must Address Deployment-Time Misalignment
Pre-deployment assessments fail to capture motivations that emerge after an AI is released. A new proposal on the AI Alignment Forum argues that dangerous motivations can develop during active deployment. This gap makes current risk analysis insufficient.
The Signal
Evaluators must now integrate deployment-time spread into their safety frameworks to prevent adversarial misalignment in future models.