Matthieu Aureille ORCID LinkedIn
Department of Computer Systems, Faculty of Information Technology, Czech Technical University in Prague; and Télécom Saint-Étienne, Jean Monnet University.
Research release · 2026
Agentic LLMs for Network Attack Detection
Can a tool-using LLM agent independently investigate a raw network capture and identify an attack?
This benchmark evaluates an LLM acting as an autonomous packet analyst, rather than a classifier over precomputed features. Given only a PCAP path and natural-language descriptions of 11 possible attacks, the agent chooses and executes its own queries, adapts to successive observations, and returns one capture-level decision.
What was tested
Autonomy
The agent receives no packet summary, extracted flow features, or analyst assistance. Its tool calls and complete investigation trajectory are retained for inspection.
Traffic conditions
Low-background captures contain only the scenario traffic and the testbed's normal idle activity. Generated-high-background captures add substantially more benign traffic, making attack evidence harder to isolate.
Matched benchmark
44 thirty-minute captures, six prompt variants, and 18 model/reasoning configurations produce 4,752 complete investigations.
Attack coverage 11 indicative classes
This mapping is for reader orientation only; these names were not shown to the LLM.
- Attack 1
- TCP ACK flood using IP fragmentation
- Attack 2
- ARP spoofing
- Attack 3
- DNS flood
- Attack 4
- FIN scan
- Attack 5
- ICMP fragmentation flood
- Attack 6
- NULL scan
- Attack 7
- TCP SYN port scan (
nmap -sS) - Attack 8
- Slowloris
- Attack 9
- UDP fragmentation flood
- Attack 10
- UDP scan
- Attack 11
- Xmas scan
Project resources
Authors
Suggested citation
If you use this benchmark, cite the canonical dataset release and reference the code repository.
M. Aureille and J. Fesl, “Agentic LLMs for Network Attack Detection,” version 1.0.0, Hugging Face dataset, 2026. [Online]. Available: https://huggingface.co/datasets/maureille/agentic-pcap
BibTeX
@misc{aureille2026agentic,
author = {Matthieu Aureille and Jan Fesl},
title = {Agentic LLMs for Network Attack Detection},
year = {2026},
howpublished = {Hugging Face dataset},
note = {Version 1.0.0},
url = {https://huggingface.co/datasets/maureille/agentic-pcap}
}Licenses
CC BY 4.0
PCAPs, dataset metadata, prompts, recorded sessions, decisions, result data, and figures.
MIT
Benchmark software and the HTML, CSS, and JavaScript used to present this site and dashboard.
Apache 2.0
The two mirrored third-party Gemma 4 chat templates retain their upstream license.