MUMBAI, India, Oct. 26 -- Intellectual Property India has published a patent application (202611113303 A) filed by Ajay Kumar Garg Engineering College on September 22, 2026, for A System And Method For Conversation-Level Detection Of Decomposed-Intent Prompt Injection In Code-Mixed Dialogue.
Inventors include Anuj Kumar; Khushi Bansal; Jhanvi Awasthi; Shruti Maurya; and Anshuman Pandey.
The application for the patent was published on October 23, 2026, under issue no. 43/2026.
Abstract: A system (100) and method (200) are disclosed for detecting a harmful intent decomposed across multiple, individually low-risk turns of a code-mixed conversation, which conventional per-turn detectors cannot detect. A detector-agnostic adapter (104) preserves a wrapped single-turn detector's per-turn path (106), while a trajectory feature extraction engine (110), using a reused multilingual encoder (108), derives risk-trajectory, semantic-drift, cumulative-convergence, hazard-topic proximity and surface-cue feature blocks (112–120) mapped by a lightweight trajectory classifier (122) to a conversation-level risk. A risk fusion module (124) combines both signals by a non-decreasing function so per-turn recall is never degraded, and a threshold module (126) outputs an explainable BLOCK/ALLOW decision (128). A benchmark module (130) validates performance. The invention attains substantially improved decomposed-intent detection at low CPU-only latency. Reference Figure:
Disclaimer: Curated by HT Syndication.
