When Do Decision Models Help LLM-Driven Network Discovery?

Network discovery, the identification of live hosts and their exposed services, is the first phase of every penetration test, and it is increasingly automated by harnesses driven by large language models (LLMs). To mitigate hallucinated or unsafe commands, recent work proposes inserting fast, non-generative decision models to validate actions prior to execution. This paper quantifies whether such validation improves discovery across three network scales (n = 10, 100, and 1,000) in a simulation calibrated on a 30-node containerized testbed. Our findings show that chunked batch dispatch recovers 92% to 100% of open ports in large-scale networks, whereas single-host dispatch collapses to under 6% due to context saturation and scan latency. Furthermore, at the measured proposal quality of the local model, decision gating increases its port recall from 75.5% to 89.0% on small networks and intercepts costly decoy tarpits, but it imposes additional latency on frontier models during early host discovery.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-18
DOI
https://doi.org/10.48550/arxiv.2611.23947
Primary Topic
Network Security and Intrusion Detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

When Do Decision Models Help LLM-Driven Network Discovery?

Ertem Esiner, Matthew Caesar, Adem Orsdemir, Rochelle Xenia Mendoza Santos et al.
arXiv (Cornell University)
Network Security and Intrusion Detection
preprint

When Do Decision Models Help LLM-Driven Network Discovery?

Ertem Esiner, Matthew Caesar, Adem Orsdemir, Rochelle Xenia Mendoza Santos, Wen Shei Ong
preprint en

Abstract

Network discovery, the identification of live hosts and their exposed services, is the first phase of every penetration test, and it is increasingly automated by harnesses driven by large language models (LLMs). To mitigate hallucinated or unsafe commands, recent work proposes inserting fast, non-generative decision models to validate actions prior to execution. This paper quantifies whether such validation improves discovery across three network scales (n = 10, 100, and 1,000) in a simulation calibrated on a 30-node containerized testbed. Our findings show that chunked batch dispatch recovers 92% to 100% of open ports in large-scale networks, whereas single-host dispatch collapses to under 6% due to context saturation and scan latency. Furthermore, at the measured proposal quality of the local model, decision gating increases its port recall from 75.5% to 89.0% on small networks and intercepts costly decoy tarpits, but it imposes additional latency on frontier models during early host discovery.

arXiv (Cornell University)
Network Security and Intrusion Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon