Agentic Small Language Models for Closed-Loop Provenance-Based Intrusion Detection
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1011-1018
DOI: 10.1609/aaaiss.v10i1.43520 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Abderrahman Elhajjout und 4 weiteren Autor:innen.
Abstract (Originalsprache)
With the adoption of the Internet of Things (IoT) in critical infrastructures, investigating security incidents across connected devices has become challenging. Graph-based intrusion detection approaches show promising detection performance. However, an anomaly alert does not explain how the surrounding activity developed, and manually following this evidence is difficult in large provenance graphs that contain many benign events and highly connected entities. Threat hunting is a proactive methodology that searches for hidden or undetected threats and reconstructs the activity surrounding an alert. Existing language-model-based approaches mainly rely on computationally expensive Large Language Models (LLMs) and model-generated graph queries, while directly presenting a complete provenance graph exceeds the reasoning and context capacity of Small Language Models (SLMs). In this paper, we propose ProvNav-SLM, an agentic threat-hunting framework that combines a Graph Neural Network (GNN) with an SLM for provenance investigation. The framework operates through three systematic phases: Detect, where the GNN identifies anomalous entities; Investigate, where an SLM agent repeatedly selects among recorded graph-navigation actions; and Verify, where a rule-based executor retrieves the selected relationships and preserves their links to the collected logs. To support SLM reasoning, the framework transforms large graph neighborhoods into small action groups organized by direction, relation, and entity type. The main novelty of ProvNav-SLM is its ability to perform closed-loop provenance threat hunting with a local SLM while separating graph-wide anomaly detection, semantic action selection, and evidence verification across dedicated components. In an exploratory evaluation on four ProvICS campaigns, the local SLM recovered more target processes on average than the evaluated graph-based rules.
Zitieren
APA
Elhajjout, A., Houda, Z. A. E., Jarir, Z., Davoust, A., & Moudoud, H. (2026). Agentic Small Language Models for Closed-Loop Provenance-Based Intrusion Detection. Proceedings of the AAAI Symposium Series, 10(1), 1011-1018. https://doi.org/10.1609/aaaiss.v10i1.43520
Vancouver
Elhajjout A, Houda ZAE, Jarir Z, Davoust A, Moudoud H. Agentic Small Language Models for Closed-Loop Provenance-Based Intrusion Detection. Proceedings of the AAAI Symposium Series. 2026;10(1):1011-1018. doi:10.1609/aaaiss.v10i1.43520
BibTeX
@article{Elhajjout2026,
author = {Elhajjout, Abderrahman and Houda, Zakaria Abou El and Jarir, Zahi and Davoust, Alan and Moudoud, Hajar},
title = {Agentic Small Language Models for Closed-Loop Provenance-Based Intrusion Detection},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {1011--1018},
doi = {10.1609/aaaiss.v10i1.43520}
}
EUVIMED