Ranjan Kathuria's 2026 research, published in the International Journal of Science and Research Archive, proposes a security-by-design framework for enterprise LLM agents on Google Cloud Platform. These agents, which perform tasks like code analysis and documentation summarization, face risks including data exfiltration, prompt injection, and over-privileged tool access. The study focuses on systems using Vertex AI with Model Context Protocol (an open standard for connecting AI assistants to external data and tools) to interact with external systems like GitHub.

The proposed architecture implements preventative infrastructure controls to establish secure boundaries before runtime. This design utilizes VPC Service Controls to create an API-level service perimeter, Private Service Connect to maintain traffic within private networks, and Identity and Access Management (IAM) Deny policies for non-bypassable guardrails. Additionally, the framework mandates that all tool credentials be stored in Secret Manager with encryption at rest and restricts agents to narrow, non-destructive Model Context Protocol tool actions.

To support observability and incident response, the architecture centralizes telemetry, including LLM call logs and tool invocations, by exporting data to security information and event management (SIEM) systems. Evaluation via a quantitative risk model shows that this integrated approach of infrastructure, identity, and monitoring controls reduces overall platform risk by approximately 91.33%. These findings demonstrate that rigorous preventative controls can significantly strengthen the security posture of enterprise AI workflows against common automated attack vectors.

Ranjan Kathuria's 2026 research, published in the International Journal of Science and Research Archive, proposes a security-by-design framework for enterprise LLM agents on Google Cloud Platform. These agents, which perform tasks like code analysis and documentation summarization, face risks including data exfiltration, prompt injection, and over-privileged tool access. The study focuses on systems using Vertex AI with Model Context Protocol (an open standard for connecting AI assistants to external data and tools) to interact with external systems like GitHub.

The proposed architecture implements preventative infrastructure controls to establish secure boundaries before runtime. This design utilizes VPC Service Controls to create an API-level service perimeter, Private Service Connect to maintain traffic within private networks, and Identity and Access Management (IAM) Deny policies for non-bypassable guardrails. Additionally, the framework mandates that all tool credentials be stored in Secret Manager with encryption at rest and restricts agents to narrow, non-destructive Model Context Protocol tool actions.

To support observability and incident response, the architecture centralizes telemetry, including LLM call logs and tool invocations, by exporting data to security information and event management (SIEM) systems. Evaluation via a quantitative risk model shows that this integrated approach of infrastructure, identity, and monitoring controls reduces overall platform risk by approximately 91.33%. These findings demonstrate that rigorous preventative controls can significantly strengthen the security posture of enterprise AI workflows against common automated attack vectors.