Controlling Artificial Intelligence Risk in Cybersecurity: A Modern Review of Governance, Protection, and the AICURE-Grid Unified Enforcement Framework

Abstract

Artificial intelligence has evolved from a narrow analytical tool into a dual-use cyber capability that supports defensive acceleration while enabling offensive automation. In defense, AI can summarize threat intelligence, triage alerts, detect anomalies, support secure code review, prioritize vulnerabilities, and enrich incident response. The same capabilities may also generate persuasive phishing, automate reconnaissance, mutate malware, exploit leaked credentials, create synthetic impersonation media, and manipulate AI agents through prompt injection or tool misuse. This review treats AI cybersecurity as an adaptive socio-technical risk rather than a conventional software defect and proposes AICURE-Grid, a layered enforcement architecture for controls before, during, and after AI-supported actions. The framework integrates governance, AI asset inventory, identity assurance, an AI gateway, policy decision logic, secure retrieval-augmented generation, data-loss prevention, model and prompt controls, sandboxed tool execution, telemetry, red-team evaluation, and incident response. It argues that model-level safety instructions are necessary but insufficient for high-impact cyber decisions, which require deterministic external controls. A measurable validation plan is provided for security, usability, traceability, and operational performance.