Kristyanto, Marco Ariano (2026) AUDY: An Autonomous LLM-Driven AI Agent for Localized Automated Infrastructure Penetration Testing. In: ICITISEE. (Submitted)
![[thumbnail of AUDY paper - draft.docx]](/style/images/fileicons/text_html.png) |
HTML
AUDY paper - draft.docx Restricted to Registered users only Download (190kB) | Request a copy |
Abstract
Abstract—The rapid expansion of IT infrastructure has dramatically expanded the cyberattack surface. Still, conventional manual penetration testing cannot keep up with continuous deployment cycles amid an ongoing cybersecurity skills gap. In this work, we present AUDY, a localized, autonomous LLM agent architecture for automated penetration testing and security reporting on infrastructure in native Bahasa Indonesia. Audy is based on the Mistral Small 3.2 24B base model and fine-tuned with Quantized Low-Rank Adaptation (QLoRA) on a domain-specific dataset of 4,934 cybersecurity samples. The agent executes within a standalone Kali Linux Docker container, as part of a decoupled reasoning-execution pipeline driven by the Model Context Protocol (MCP) and regulated by a two-layer safety control architecture that combines system instructions (SOUL.md) with a deterministic target-scope guardrail (_guard()). In the experiments, we find that QLoRA fine-tuning achieves a test token accuracy of 95.31% and a test loss of 0.1604 without overfitting. AUDY was successfully deployed in live tests against authorized institutional infrastructure, where it executed automated reconnaissance workflows strictly intercepting out-of-scope targets. Hardware evaluation on a single NVIDIA RTX PRO 6000 Blackwell MIG instance shows that context window scaling up to 64K tokens remains within safe VRAM margins (94.96% usage). This study shows the promise of privacy-preserving, local LLM agents in democratizing cybersecurity assessments for low-resource languages.
| Item Type: | Conference or Workshop Item (Paper) |
|---|---|
| Subjects: | T Technology > T Technology (General) |
| Divisions: | Faculty of Engineering > Department of Information Technology |
| Depositing User: | MARCO ARIANO KRISTYANTO |
| Date Deposited: | 30 Sep 2026 06:34 |
| Last Modified: | 30 Sep 2026 06:34 |
| URI: | http://repository.ubaya.ac.id/id/eprint/51309 |
Actions (login required)
 |
View Item |
