A new audit on AI model maintenance warns that “machine unlearning” can silently degrade performance on rare diagnoses and proposes a safety framework to detect and mitigate spillover risks. In another deployment-oriented clinical AI update, an international study showed adult-trained chest X-ray AI models can adapt to newborn images via continual learning across 30 neonatal intensive care units without patient data sharing—aiming to make model updates feasible at scale. Together, the developments show both the operational risk (performance drift) and the practical path (continual adaptation) toward safer, more reliable clinical AI rollouts.