Zscaler rolled out a full zero trust stack for agentic AI at its Zenith Live 2026 conference in Las Vegas. The platform sits on its existing Zero Trust Exchange and adds four key components. First, the Zscaler AI Broker locks down how AI agents talk to data (Model Context Protocol) and to each other (Agent-to-Agent). Second, Endpoint AI Security watches browsers, plugins, extensions and local AI tools on employee devices—areas traditional endpoint products often miss. Third, the AI Access Graph, built on Zscaler’s Symmetry Systems acquisition, maps every identity, app and data source so you can enforce policies and track data flows in real time. Finally, AI Protect now includes an AI asset manager that discovers agents and embedded AI across SaaS and internet traffic, scans code bases for risks and offers prompt extraction plus compliance checks for over 250 GenAI apps. Beyond those core tools, Zscaler is adding AI red-teaming against MCP servers, a prompt-hardening service and heat maps for compliance. The expanded controls cover multi-turn conversations in Anthropic and OpenAI APIs, and intent-based guardrails that stop unauthorized actions before they begin. Jay Chaudhry, Zscaler’s CEO, pointed to the millions of autonomous agents operating at machine speed—an environment traditional security wasn’t built for. Security vendors see a booming market. Dell’Oro Group forecasts AI systems security will swell from virtually zero last year to $8 billion by 2030, with nearly 60 providers already in play. Most focus on model security, runtime guardrails, AI posture management or red-teaming. Analysts stress zero trust must extend to agents themselves. Mauricio Sanchez of Dell’Oro says each agent needs a unique identity, scoped permissions and continuous monitoring—no inherited trust just because a user launched it. Michela Menting at ABI Research warns that without these controls, bad actors or simple misconfigurations could let agents move laterally, escalate privileges and hit critical systems in seconds.