New research report accepted for submission to ICAART 2026

The research report “JewelCVSS: A Domain-Tuned LLM for Automated Vulnerability Scoring”, carried out by colleagues Stefano Valcada, Antonio Maci, Roberto Lorusso, Alessandro Santorsola and Pietro Spalluto as part of the research activities envisaged by the Program Contract “Cybersecurity and SOC Product Suites”, was accepted for presentation at the “ICAART 2026 – 18th International Conference on Agents an Artificial Intelligence”, to be held in Marbella, Spain, from 5 to 7 March 2026, and will be published later in the conference proceedings.

The research proposes a Large Language Model (LLM)-based solution that can preliminarily, but with high accuracy, estimate the Vulnerability Score, i.e., the degree of severity associated with a vulnerability according to the Common Vulnerability Scoring System (CVSS), as soon as it is published in the Common Vulnerability and Exposures (CVE) databases along with its description.

The model used is appropriately fine-tuned, that is, customized through targeted training on specific data and cases from the cybersecurity domain, thus improving understanding of the technical language and typical vulnerability dynamics. This customization allows the model to provide more reliable assessments that are consistent with the standards adopted by human analysts.

The solution is therefore able to offer an immediate indication of the impact of a vulnerability to all operators in the sector as soon as it is made public, well before the official assignment of the Vulnerability Score. This increases the timeliness of remedial interventions in critical cases and helps reduce the time window during which attackers can effectively exploit vulnerability.

The activity was carried out exclusively by BV TECH personnel within the CyberLab of Grottaglie and Rutigliano, demonstrating the skills and abilities developed in the use of Generative AI applied to the cybersecurity domain.