Adversarial attacks can target AI models in industrial control environments, alongside risks such as data poisoning, model tampering, and compromised supply chains. These threats have cyber-physical security implications for AI-connected safety systems, requiring network segmentation and strict access control for inference engines within OT networks.