Autonomous agentic artificial intelligence (AI) systems increasingly operate as decision-making entities capable of selecting tools, invoking services, processing information, and executing actions with limited human intervention. This autonomy creates an access-control problem that cannot be adequately addressed through static permission models because the security significance of an action depends on contextual variables such as requested resource, operational state, transaction sensitivity, delegation history, environmental conditions, and estimated risk. This paper proposes a dynamic risk-based access-control architecture for autonomous agentic AI systems that combines identity verification, contextual risk assessment, policy evaluation, least-privilege enforcement, adaptive authorization, and continuous security monitoring. The proposed methodology conceptualizes authorization as a dynamic decision function rather than a one-time permission assignment. The architecture introduces a multi-stage policy-enforcement pipeline in which an agent's identity and delegated authority are evaluated together with action context and risk indicators before access is granted, constrained, escalated, or denied. The research further develops a security-evaluation framework covering authorization accuracy, privilege containment, policy consistency, decision latency, adaptability, and resistance to privilege escalation. The risk-adaptive authorization concept is particularly relevant because recent work demonstrates that contextual authorization can combine workload identity, policy evaluation, risk scoring, constrained tokens, and differentiated approval levels for autonomous agents (Bhushan, Pappu, Jadav, and Jaiswal, 2026). Drawing on optimization-oriented and data-driven system research, the proposed framework emphasizes continuous recalibration rather than fixed authorization thresholds. Analytical findings indicate that dynamic authorization can improve security resilience and reduce unnecessary privilege exposure while introducing computational and governance trade-offs. The study contributes an integrated conceptual architecture and evaluation model for implementing risk-sensitive access control in autonomous AI environments.