As AI agents grow more autonomous and capable, they introduce novel security challenges that traditional cybersecurity approaches may not fully address. In a new episode of Practical AI's "Fully Connected" series, hosts Dan Whitenack and Chris Benson dive into Anthropic's Zero Trust for AI Agents security framework, designed to help organizations safely deploy increasingly sophisticated agentic systems without compromising security posture. The discussion unpacks how traditional Zero Trust principles—originally developed for human-managed networks—must evolve when applied to autonomous AI agents that operate independently and make real-world decisions. Whitenack and Benson examine the specific security risks that emerge when agents have greater autonomy, including potential misuse, unintended actions, and cascading failures across integrated systems. They break down practical security controls organizations can implement immediately, while exploring how cybersecurity architecture itself must adapt for an age where non-human actors execute sensitive operations. The episode highlights the growing urgency of this challenge as enterprises increasingly adopt agentic AI for high-impact environments. Prediction Guard, a self-hosted AI control plane for agent deployment, was featured as a relevant infrastructure solution in this space, while links to OWASP's GenAI Project underscore the broader industry push toward standardized security guidelines for generative AI systems.