MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641076110 A) filed by Rashmi T V; Kanchana R; Dr. I Manimozhi; Revathi K; Triveni N; Madhu Shree R; and Divya Shree N on June 19, 2026, for Securing Agentic Ai Systems: Security Risks In Autonomous Ai Agents, Prompt Injection Prevention, And Trust Management Frameworks.
Inventors include Kanchana R; Dr. I Manimozhi; Revathi K; Triveni N; Madhu Shree R; and Divya Shree N.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: Agentic Artificial Intelligence (AI) systems, powered by Large Language Models (LLMs), are increasingly deployed as autonomous agents that plan, reason, and execute actions across digital environments through tools, APIs, and persistent memory. This autonomy fundamentally changes the security calculus: unlike predictive models, an agent can take consequential real-world actions, so a single manipulated instruction can cause data exfiltration, unauthorized transactions, or self-corruption of memory. This paper presents a comprehensive survey of security risks in agentic AI systems and a formal threat model that decomposes the agent into an explicit attack surface. We propose the Secure Agentic AI Framework (SA³F), a defense-in-depth architecture that integrates input sanitization, a prompt firewall, constrained reasoning, a tool-execution sandbox, memory protection, an explainable trust-management engine, and human-in-the-loop governance. We formalize a trust-scoring and risk-gating model, specify the end-to-end secureexecution algorithm, and present a worked case study in which an indirect prompt-injection attack delivered through retrieved web content is neutralized. A comparative analysis against standard LLMs, rule-based guardrails, and retrieval-augmented agents highlights gaps in current agentsecurity research, and we propose a layered evaluation methodology and threat-coverage matrix to guide empirical validation. The contribution is an integrated, explainable, and deployable blueprint rather than empirically validated outcomes
Disclaimer: Curated by HT Syndication.
