MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202611090918 A) filed by Manipal University Jaipur on July 24, 2026, for System And Method For A Multi-Layered Defense Against Prompt Injection Attacks In Large Language Models.
Inventors include Mritunjay Singh; Navyaa Dwivedi; Sanyam Kathed; Mridul Maheshwari; and Dr. Akhilesh Kumar Sharma.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: The present invention relates to a computer-implemented multi-layered defense system (100) and method for protecting Large Language Models (110) against prompt injection attacks. The system comprises an input interface (101) for receiving prompt inputs, a preprocessing engine (102) for context-aware normalization of prompt data, and a Dynamic Semantic State Machine (103) for analysing semantic intent and detecting contextual manipulation. A feature extraction engine (104) generates multidimensional features that are processed by an ensemble risk assessment engine (105) to determine a prompt risk score. The system further includes an invariant validator (106), a constitutional behaviour validator (107), an adaptive sanitization engine (108), a secure execution controller (109), and a feedback and adaptive learning engine (111) for continuously updating detection models. The invention provides a defence-in-depth architecture that accurately detects, mitigates and prevents prompt injection and jailbreak attacks while preserving legitimate user intent and enabling secure, reliable and real-time deployment of Large Language Models.
Disclaimer: Curated by HT Syndication.
