A Zero‑Trust Compliance Architecture for LLM‑Integrated Pharmaceutical IT Systems: Securing AI‑Assisted Workflows with Data Integrity and Regulatory Controls

Abstract

LLMs are being used in regulated pharmaceutical IT workflows for standard operating procedure (SOP) generation, deviation triage and Corrective and Preventive Action (CAPA) writing, but this is complicated by stochastic responses and susceptibility to prompt injection along with issues of data integrity, role-based access control and auditability. This paper proposes the Zero‑Trust Compliance Architecture (ZT-LLM-COMPASS) for this zero-trust setting, where every model invocation, context retrieval, tool invocation, and artifact generation is treated as a potential attacker payload. ZT-LLM-COMPASS incorporates four LLM Security Planes: fine-grained authorization via Identity & Policy Enforcement Plane, prompt firewalls and controlled retrieval via Prompt & Context Security Plane, least-privileged permissions on function calls via Tool Sandbox, and tamper-obvious audit packages via Evidence Plane. Human-in-the-loop verification gates ensure that the LLM can only propose and never commit regulated records. Adversarial evaluation demonstrates that the design is strong against injection attacks, leakage, and auditing challenges, enabling high AI productivity while retaining regulatory compliance.