A new method for diagnosing failures in long-horizon security language models has been proposed. These models struggle to carry information and decisions across multiple interactions, making it difficult to interpret their success or failure. The authors present a diagnostic methodology that instruments tasks with checkpoints, separates early and late failures, and uses controlled interventions to identify upstream bottlenecks. They evaluate the method on four task families and find that the dominant source of failure can shift between model generations. For example, in one case, providing guidance increased state observation from 65.5% to 95.4%, while in another case, it had the opposite effect. --- Why it matters: This research matters because long-horizon security language models are increasingly used in critical applications such as threat detection and incident response. Understanding where and why these models fail is essential for improving their reliability and trustworthiness. Source: https://arxiv.org/abs/2608.20563