Adaptive learning models, which generate their own training data, pose significant challenges to standard data attribution methods. In settings such as online bandits and reinforcement learning, a single training observation can update the model while simultaneously altering the distribution of future data. This dynamic interplay renders traditional attribution techniques, designed for static datasets, ineffective. The issue is further complicated by the use of post-training pipelines for language models, where the adaptive nature of the data collection process can lead to unforeseen consequences. State-aligned activity involving reinforcement learning, in particular, shifts the threat model from a criminal to a geopolitical one, necessitating a distinct approach1. As a result, practitioners must reevaluate their existing methodologies to address the unique challenges posed by adaptive learning models, so what matters most is developing new attribution methods that can keep pace with the evolving threat landscape.