Option D is correct because the AI application is taking high-impact remediation actions without adequate human authorization. Human-in-the-loop control requires a qualified person to review and approve consequential actions, such as stopping servers or changing firewall exposure, before execution. The system may still detect misconfigurations, rank risks, and recommend fixes, but it should not autonomously disrupt development or alter network access during a demonstration. Option A restores service temporarily but leaves the unsafe decision path unchanged. Option B removes useful monitoring and does not correct the excessive automation. Option C might close improperly opened ports, yet it addresses only one symptom and does not prevent the AI from repeating other damaging actions. The stronger design is to roll back the unauthorized changes, restrict the agent’s permissions, require approval for disruptive operations, and maintain an auditable record of recommendations and approvals. The NIST AI Risk Management Framework calls for appropriate human-oversight processes to be defined, assessed, and documented, supporting this governance pattern.

================