Intellectual Property India Publishes Patent Application for An Artificial-Intelligence System And Method That Keeps An Ai Assistant Inside The Permission Its User Actually Gave And That Checks Every Tool

MUMBAI, India, Sept. 14 -- Intellectual Property India has published a patent application (202641104360 A) filed by Rajalakshmi Engineering College on August 31, 2026, for An Artificial-Intelligence System And Method That Keeps An Ai Assistant Inside The Permission Its User Actually Gave And That Checks Every Tool.

Inventors include Dr. Suresh Kumar S; and Dr Sudalaimuthu T.

The application for the patent was published on September 11, 2026, under issue no. 37/2026.

Abstract: AN ARTIFICIAL-INTELLIGENCE SYSTEM AND METHOD THAT KEEPS AN AI ASSISTANT INSIDE THE PERMISSION ITS USER ACTUALLY GAVE, AND THAT CHECKS EVERY TOOL BEFORE THE TOOL IS USED Al software agents now carry out jobs on their own for a user, using outside tools and handing parts of the work to other agents. Such an agent can be fooled by text that did not come from the user - hidden inside a document it reads, inside a tampered tool description, or inside a message from another agent - and it then uses the permission it genuinely holds to do something the user never asked for. The present application addresses this problem, known as the confused deputy problem, which existing protections do not solve because they check only the credential an agent carries and never check the request against what the user actually wanted. A secure agent delegation system (100) and method (200) are described. An intent extraction module (102) reads the user's request and writes down an intention note containing a goal, an allowed area of work, and a never-do list. A token issuance module (104) makes a permission slip that is cryptographically locked to that note, and a delegation graph manager (106) passes the slip through several hand-overs, making the allowed area smaller each time. A toolchain attestation module (108) lets a tool into the trusted circle only after checking the signature and the content fingerprint of its manifest and after deciding that the tool's declared purpose matches the note. A runtime monitor and policy enforcement module (110) checks every tool request against the allowed area and against a behaviour agreement built from the note, and an oversight module (112) sends any request above a set risk level to a human for approval. In this way, a tool request that the agent's credential would allow but that does not match what the user (101) wanted is stopped before it can run, and any change made to a tool description after approval is detected.

Disclaimer: Curated by HT Syndication.