AI agents can answer questions, search documents, call tools, remember information, and take action-but what happens when an untrusted prompt, poisoned document, compromised memory, or unsafe tool request changes what they do?
AI Agent Security with Python and MCP turns this complex subject into a practical, step-by-step learning journey. You do not need previous experience in AI security, prompt injection, red teaming, RAG poisoning, MCP security, or multi-agent defence. Basic Python familiarity is helpful, but every major concept is introduced clearly before you apply it.
Instead of overwhelming you with theory, the book guides you through safe, controlled laboratories built with fictional systems, synthetic data, mock credentials, and isolated services. You will attack deliberately vulnerable AI agents, observe what fails, implement layered defences, and verify that the same attacks no longer succeed.
Mistakes are treated as useful evidence. Each blocked tool call, quarantined document, rejected memory record, and failed release gate becomes a small, measurable win that builds your confidence.
KEY FEATURES
- A beginner-accessible Attack-Defend-Verify methodology
- Six connected, hands-on security projects
- Practical Python examples presented in manageable steps
- Safe red-team exercises using synthetic data and mock services
- Reusable threat models, policies, tests, metrics, and recovery playbooks
- Coverage of prompt injection, poisoned RAG, MCP tools, memory, multi-agent systems, observability, and CI/CD security gates
WHAT YOU WILL LEARN
- Threat-model AI-agent applications and identify critical trust boundaries
- Reproduce direct and indirect prompt-injection attacks safely
- Protect RAG pipelines with source verification, provenance, quarantine, retrieval filtering, and citation checks
- Secure MCP tools with authentication, authorisation, least privilege, approval, idempotency, revocation, and audit trails
- Defend persistent memory and shared multi-agent state from poisoning and contamination
- Build an Agent Security Gateway with policy-as-code, risk controls, sandboxing, and kill switches
- Detect incidents, contain compromised components, recover trusted state, and convert failures into regression tests
- Build AgentShield, a red-team CI platform that blocks unsafe releases
WHO THIS BOOK IS FOR
This book is for Python developers, AI and machine-learning engineers, MCP and FastAPI developers, AppSec and DevSecOps professionals, penetration testers, security architects, technical leads, students, and self-learners who want a practical introduction to AI-agent security.
TABLE OF CONTENTS
Building a Safe Agent Security Laboratory
Threat-Modelling Agentic Systems
Prompt-Injection Attacks and Defences
Securing RAG Against Poisoned Knowledge
Building the Secure MCP Refund Assistant
Defending Agent Memory and Multi-Agent Workflows
Building the Agent Security Gateway
Security Observability and Incident Response
Building the AgentShield Red-Team CI Platform
Running the Integrated Security Campaign
Build, attack, defend, test, and operate AI agents with greater confidence. Start your practical AI-security journey today.
