Zero Trust in the Age of AI: Why the Classic Model Isn’t Enough Anymore

Classic Zero Trust was designed around human identities, but AI agents now outnumber human users in many enterprises — sometimes 10:1 or higher. AI agents behave fundamentally differently: they make thousands of API calls autonomously, traverse multiple systems in a single workflow, inherit broad permissions, and can be manipulated via prompt injection. This creates security gaps that traditional Zero Trust frameworks weren't built to address. Key adaptations include giving every AI agent its own scoped workload identity with short-lived credentials, implementing just-in-time permission grants for agentic workflows, sanitizing agent inputs to prevent prompt injection, building agent-specific behavioral baselines separate from human UEBA, and adding human-in-the-loop checkpoints for irreversible high-stakes actions. A practical roadmap for CISOs covers immediate steps (inventory AI tools, audit service accounts), year-long initiatives (workload identity management, agent logging, prompt injection assessments), and ongoing practices (red team AI scenarios, tracking NIST SP 800-207 updates).