A Runtime Safety Copilot for AI-Native O-RAN: Predictive Verification and Fail-Safe Enforcement in Near-RT RIC Control Loops

Sixth-generation (6G) mobile networks are expected to be AI-native by design, with intelligence embedded directly into the Radio Access Network (RAN) control loop through architectures such as O-RAN, near-real-time RAN Intelligent Controllers (near-RT RICs), and AI-driven xApps and rApps. While considerable research has focused on performance optimisation and security of AI-enabled RAN, the problem of engineering explicit runtime safety guarantees for AI-driven control decisions remains largely unaddressed. In particular, current AI-RAN solutions lack systematic mechanisms to prevent unsafe actions during critical operations such as handover management, beamforming adaptation, and network slicing control, where erroneous or unstable decisions can lead to service disruption, cascading failures, or violation of service-level agreements. This work introduces the Safety Copilot: an independent, runtime-enforced supervisory layer that continuously intercepts and verifies AI-generated control actions before they are executed in the live control plane. The framework combines runtime safety verification, constraint-based control enforcement, and control-plane reliability engineering to ensure that AI-generated actions comply with predefined safety invariants and operational limits. Unlike conventional offline model validation or post-hoc monitoring, the proposed approach performs both pre-deployment and in-operation verification, enabling fail-safe intervention, decision rollback, or graceful degradation when unsafe behaviours are detected. Evaluated over <inline-formula> <tex-math notation="LaTeX">$T = 12{,}000$ </tex-math></inline-formula> decision epochs across four competing control strategies, the Safety Copilot intercepts 32.27% of unsafe AI-proposed actions while preserving throughput comparable to the performance-driven baseline, reducing handover instability by 46% and ping-pong rate 72% relative to the baseline AI. Verification latency remains within microsecond-scale bounds (median <inline-formula> <tex-math notation="LaTeX">$= 9~\mu $ </tex-math></inline-formula>s, p<inline-formula> <tex-math notation="LaTeX">$95= 13~\mu $ </tex-math></inline-formula>s), confirming compatibility with near-RT RIC control loop timing budgets. By separating safety enforcement from both performance optimisation and security mechanisms, the proposed framework establishes a proof-of-concept foundation toward trustworthy, resilient, and standards-aligned AI-native 6G networks, with testbed-level validation identified as a priority for future work.