MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641122374 A) filed by Dr. Theresa Cenate C. F; Durga Aditya Pandey; Anirudh Sarkar; and Vinay Krishnan on October 13, 2026, for System And Method For Detecting And Mitigating Obfuscated Jailbreak Attacks On Transformer Language Models By Layer-Wise Analysis Of Internal Activations At A De-Noising Horizon.
Inventors include Dr. Theresa Cenate C. F; Durga Aditya Pandey; Anirudh Sarkar; and Vinay Krishnan.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: ABSTRACT SYSTEM AND METHOD FOR DETECTING AND MITIGATING OBFUSCATED JAILBREAK ATTACKS ON TRANSFORMER LANGUAGE MODELS BY LAYERWISE ANALYSIS OF INTERNAL ACTIVATIONS AT A DE-NOISING HORIZON A system and method detect obfuscated jailbreak prompts from the internal activations of a transformer language model during its forward pass. Hidden states captured at monitored layers are projected through the output head of the model to obtain a refusal-concept mass per layer, from which an elevation ratio and an onset layer (a De-Noising Horizon) are derived. Optionally, a Shannon effective rank of pooled activations, whose contraction signals templated obfuscation, and a layer-wise entropy trajectory are fused into an obfuscation-risk indicator. When the indicator meets a criterion, a protective action is executed before any responsive token is released, including withholding, routing, or guarded steering that projects a steering vector away from a generation subspace, bounds its norm and supervises coherence. The approach is wrapper-agnostic and reuses computation of the forward pass.
Disclaimer: Curated by HT Syndication.
