A Hybrid Regex-Semantic Scoring Framework for Robust Jailbreak Prevention in Large Language Models
International Journal of Sciences and Innovation Engineering · 12.06.2026 · 3 (6) · 728-743
DOI: 10.70849/ijsci03062672325 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Sciences and Innovation Engineering, online erschienen am 12.06.2026. Verfasst von D. Anuradha.
Abstract (Originalsprache)
Large Language Models (LLMs) are becoming an essential part of digital systems, so protecting their interactions from abuse is becoming more and more important. In order to stop phishing attempts, adversarial prompt injections, and the disclosure of private data, this project offers a general-purpose jailbreak detection and safety framework. To identify both direct and covert dangerous inputs, the system combines semantic similarity analysis with rule-based filtering utilizing regular expressions. In real time, it examines user prompts and LLM outputs, preventing any that semantically resemble carefully chosen jailbreak samples or fit known exploit patterns. To facilitate auditing and monitoring, all interactions are also recorded with timestamps and safety status. This solution provides a scalable and transparent method for guaranteeing safe and responsible LLM deployment across a variety of application domains by fusing conventional heuristic detection with AI-powered semantic comprehension.
Zitieren
APA
Anuradha, D. (2026). A Hybrid Regex-Semantic Scoring Framework for Robust Jailbreak Prevention in Large Language Models. International Journal of Sciences and Innovation Engineering, 3(6), 728-743. https://doi.org/10.70849/ijsci03062672325
Vancouver
Anuradha D. A Hybrid Regex-Semantic Scoring Framework for Robust Jailbreak Prevention in Large Language Models. International Journal of Sciences and Innovation Engineering. 2026;3(6):728-743. doi:10.70849/ijsci03062672325
BibTeX
@article{Anuradha2026,
author = {Anuradha, D.},
title = {A Hybrid Regex-Semantic Scoring Framework for Robust Jailbreak Prevention in Large Language Models},
journal = {International Journal of Sciences and Innovation Engineering},
year = {2026},
volume = {3},
number = {6},
pages = {728--743},
doi = {10.70849/ijsci03062672325}
}
EUVIMED