Intent-graph agent raises automated penetration-test success in CTF benchmark

A preprint describes Intentest, an automated penetration-testing agent that recorded an 88.2% overall success rate across real capture-the-flag challenges and 75.0% on hard tasks. Those results were approximately 44 and 50 percentage points above the baseline, respectively. The system stores verified network states as immutable facts in a persistent directed graph, then constrains subsequent exploration to intents supported by those facts. This is a benchmark result rather than evidence of performance on production networks, and the paper has not established that the gains transfer beyond its challenge set.