AI-Mediated Interpretive Integrity Attacks (AMIIA): A Formal Threat Model for Silent Policy Drift in AI-Mediated Workflows

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI-mediated workflows increasingly summarize policies, interpret compliance requirements, draft standard operating procedures (SOPs), triage alerts, and support risk decisions. This creates a security-relevant interpretive layer between canonical governance artifacts and operational practice. This paper defines AI-Mediated Interpretive Integrity Attacks (AMIIA): adversarial campaigns or techniques that influence external information environments so that AI systems produce materially divergent interpretations of policy, law, control requirements, or risk criteria. The attack succeeds when personnel operationalize those interpretations without adequate validation against canonical authority, thereby weakening the effective control posture without an unauthorized technical modification. The paper distinguishes AMIIA from hallucination, misinformation, prompt injection, retrieval-augmented generation (RAG) poisoning, and model poisoning; introduces five drift surfaces and five drift primitives; specifies an attack chain, adversary model, protected assets, invariants, a scoring model, evidentiary thresholds, and a control architecture; and proposes testable claims for empirical evaluation. The central finding is that policy interpretation functions as a semantic control plane and must be governed as a controlled configuration item.