Optimizing Agentic Architectures for Cybersecurity Tasks with Trace

In this work, we introduce a novel agentic workflow that leverages Trace, a computational graph-based framework that analyzes execution traces via Directed Acyclic Graphs (DAGs) to systematically refine LLM reasoning in cybersecurity tasks. By structuring execution as a graph traversal problem, our approach enhances the model’s ability to iteratively generate, analyze, and optimize its code-based solutions, improving both reasoning depth and task success rates. We implement CyTrace on a subset of CTF tasks derived from the CyBench benchmark, covering domains such as cryptography, reverse engineering, steganography, and exploit development. Our proposed approach solves 10 tasks, achieving a $25\%$ solved rate, compared to $17.5\%$ from the base model alone, and outperforming o3-mini $(22.5\%)$.