Distributed autonomous AI systems composed of interacting intelligent agents are increasingly deployed across cloud, edge, and cyber-physical infrastructures. While these systems enable scalable decision-making and automation, they also introduce a rapidly expanding agentic attack surface consisting of vulnerabilities across model pipelines, inter-agent communication, and decision orchestration layers. Traditional MLOps frameworks focus primarily on operational efficiency and model lifecycle management but lack mechanisms to defend against adversarial machine learning attacks and systemic AI manipulation. This paper introduces an Adversarial MLOps framework designed to protect the agentic attack surface in distributed autonomous AI ecosystems. The proposed framework integrates adversarial threat modeling, secure model lifecycle management, continuous adversarial testing, and autonomous monitoring of agent behavior. A formal model for adversarial robustness in distributed AI pipelines is proposed, alongside a multi-layer architecture for securing agentic AI infrastructures. Experimental evaluations demonstrate improved resilience against adversarial attacks, reduced attack success rates, and enhanced system observability. The results highlight the importance of embedding adversarial defense mechanisms directly into AI operational pipelines to ensure trustworthy and resilient autonomous AI systems.