AI agents no longer just answer questions. They book travel, query databases, move money, write and ship code, and coordinate with other agents to get the job done, often with no human checking each step. That shift, from generating text to taking action, has quietly created a new attack surface, and most security teams are defending it with tools built for software that never had a mind of its own. This book is the practical playbook for closing that gap. Written for security engineers, AI/ML engineers, architects, and technical leaders who are building or securing agentic systems right now, it moves systematically from first principles to production-grade defense: how agents reason and act, where trust boundaries actually sit, and how each one fails. Inside, you'll work through the OWASP Top 10 for Agentic Applications and MITRE ATLAS, the two frameworks shaping how the industry classifies agentic risk, and apply the Cloud Security Alliance's MAESTRO model to multi-agent systems where a single compromised peer can quietly poison an entire pipeline. You'll dig into prompt injection and goal hijacking, the tool and MCP supply chain, memory and RAG poisoning that can sit dormant for weeks before it pays off, and the governance, identity, and kill-switch architecture that turns "a human is watching" from a comforting assumption into an actual control. Every chapter is grounded in realistic, detailed case studies, a bank support agent, a logistics red-team finding, a three-agent content pipeline that published a false claim about a competitor, and closes with a hands-on exercise you can run against your own systems, not a hypothetical one. By the end, you won't just understand agentic AI risk in the abstract: you'll have a working risk register, a tested incident response plan, and a concrete plan for maturing your organization's governance from ad hoc to genuinely defensible. Agentic AI is moving fast. This book is for the people responsible for making sure it doesn't move faster than the controls meant to contain it.
