Security-by-design for large language model platforms in google cloud platform: preventative controls for vertex AI agents with controlled external tool access

Large language model platforms are increasingly integrated into enterprise workflows, where internal artificial intelligence agents assist with tasks such as reviewing digital artifacts, summarizing technical content, and analyzing code, tickets, or documentation; code review with GitHub is one representative example of these patterns. While such systems improve productivity, they introduce new risks involving data exfiltration, over-privileged tool use, prompt injection, secret exposure, incomplete logging, and unauthorized automated actions0.
This research addresses the problem of securing an internal platform for large language model-based agents built on Google Cloud Platform using Vertex AI as the model layer and a Model Context Protocol style integration for interacting with external tools such as source control or issue-tracking systems. Following a secure-by-design methodology, the paper proposes a preventative security architecture that applies hard infrastructure, networking, identity, and monitoring boundaries before runtime interactions occur. The proposed design uses VPC Service Controls to place Vertex AI and related Google-managed services inside an API-level service perimeter that reduces data-exfiltration risk, combines Private Service Connect interfaces and egress proxies to keep agent traffic on controlled private paths, applies Identity and Access Management Deny policies to enforce non-bypassable guardrails on sensitive cloud operations, stores all tool credentials in Secret Manager with encryption at rest, and constrains agent behavior through narrowly scoped Model Context Protocol tools that expose only non-destructive actions to external systems.
In addition, the architecture centralizes observability by enabling detailed audit, access, and trace logging for large language model calls, network flows, and tool invocations, exporting this telemetry to a security information and event management platform to support detection, response, and quantitative risk assessment. The design is evaluated using a quantitative risk formula based on likelihood and impact, and the results show that the proposed architecture reduces modeled platform risk by approximately 91.33%, indicating that preventative infrastructure, identity, and monitoring controls can materially improve the security posture of enterprise large language model systems