LLM coding agents like Claude Code and OpenAI Codex can drive penetration-testing tools, but they reason about an engagement as a flat list of findings in a context window. The authors argue this misses the way attackers actually think: in graphs, where isolated low-severity issues connect into exploit chains.

The paper, posted on arXiv, proposes a Neo4j-backed approach that stores findings as a graph, letting the LLM reason over relationships between vulnerabilities. Instead of treating each finding independently, the agent can traverse connections to identify multi-step attack paths.

This shift from flat lists to graph-structured reasoning could improve how autonomous agents plan and execute penetration tests. The abstract is brief, but the core idea is clear: representing the engagement as a graph gives the model a more accurate model of the attack surface.