A Control Plane Architecture for Secure and Governable AI in Regulated Financial Systems

Abstract

Artificial Intelligence adoption within regulated financial institutions has outpaced the security frameworks designed to govern it. Existing controls were built around network perimeters, static role assignments, and API-layer enforcement and were never designed for the behavioral unpredictability of large language models (LLMs) and autonomous AI agents. This paper presents a Control Plane Architecture for Secure and Governable AI (CP-SGAI): a purpose-built governance architecture that treats AI inference as a first-class security event requiring identity attestation, policy-bounded prompt execution, and ephemeral privilege scoping, and structured observability. Grounded in field experience operating multi-account AWS environments at a large financial institution, the framework addresses practical gaps that purely theoretical governance models miss, including prompt injection at the enterprise boundary, cross-account data egress through model responses, and the absence of prompt-response lineage in existing SIEM and audit toolchains. Three original constructs are introduced: an AI Interaction Identity (AII) attestation model, an Ephemeral Prompt Security Context (EPSC) lifecycle, and an AI Observability Schema (AOS) aligned with financial regulatory audit requirements.