- Neu
E-Book, Englisch, 310 Seiten
Reihe: Packt Publishing
Vos Enterprise AI Agent Security
1. Auflage 2026
ISBN: 978-1-80882-864-5
Verlag: Packt Publishing
Format: EPUB
Kopierschutz: 0 - No protection
A Practical Guide to MCP Security, Prompt Injection Defense, RAG Security, and Agentic AI Governance
E-Book, Englisch, 310 Seiten
Reihe: Packt Publishing
ISBN: 978-1-80882-864-5
Verlag: Packt Publishing
Format: EPUB
Kopierschutz: 0 - No protection
AI agents can read sensitive data, call tools and APIs, write code, change system state, and act on behalf of users. Each action needs explicit authority and enforceable limits.
Securing Enterprise AI Agents is a practical field guide to bounded AI autonomy. Learn to classify agent risk, assign ownership and approval paths, and design a seven-layer reference architecture. Secure Model Context Protocol (MCP) servers and tool calls, apply delegated identity and least privilege, and govern retrieval augmented generation (RAG) and enterprise knowledge.
The book connects evaluations, observability and reliability controls with secure coding-agent workflows, policy as code, incident response and continuous AgentSecOps. Worked enterprise scenarios, exercises, illustrative code and a production-readiness checklist help engineering, security, risk and audit teams turn the guidance into controls they can assess.
By the end, you will be able to define boundaries for agent actions, collect audit evidence, plan for failures and assess whether the controls justify production deployment.
Autoren/Hrsg.
Weitere Infos & Material
Enterprise AI Agent Security
A Practical Guide to MCP Security, Prompt Injection Defense, RAG Security, and Agentic AI Governance
Thomas De Vos
© 2026 Thomas De Vos
Director, HCLTech AI Labs
Global Financial Services AI Lead
For Marilin
For your patience through the late nights, your judgement when mine ran out, and your love throughout. You have been my inspiration from the start, and you still are.
In memory of my father, Flor De Vos
Truly missed and certainly not forgotten. I only wish you could have proofread these pages. With your guidance, this book would be an absolute best-seller. I miss you.
Preface
What this book is
Enterprise AI agents are moving from demos into workflows that touch code, data, tools, tickets, documents, APIs, and customer operations. That changes the security problem. The old model of reviewing prompts, logging chatbot sessions, and writing acceptable-use policies is not enough.
This book shows how to secure AI agents as operational systems. It introduces bounded AI autonomy, AgentSecOps, MCP security, RAG governance, identity controls, human approval patterns, audit evidence, and regulatory readiness as one connected discipline. The message is simple: enterprises do not need fully autonomous agents. They need bounded agents that can act, prove what they did, stay inside policy, and fail safely.
The future of enterprise AI is not maximum autonomy. It is bounded autonomy with security, governance, and evidence built into the operating model.
I have been building AI systems for about eleven years and shipping software for about twenty-five. Most of that work has been in financial services, where if your system makes a mistake, somebody will eventually file a complaint or a fine. That bias shows up in the examples. Banking and insurance companies do not have a monopoly on caring about safety, but they do have a habit of caring about it earlier than most.
The book tries to be useful to four kinds of reader. The engineer who has built an agent demo and is now being asked to put it in front of customers. The architect who has been handed a stack of agent proposals and needs a way to triage them. The security or risk lead who has been asked to sign off on something they cannot fully see. The engineering director who has to decide where to spend the next quarter of platform investment.
If you are any of those people, you can read it cover to cover. If you are short on time, the chapter summaries, the unified AgentSecOps architecture in Chapter 20, and the production readiness checklist in Chapter 21 will get you most of the way.
Who this book is for, and how to scale the burden
The mental model that organises this book in one sentence: an AI agent is a new, highly privileged employee. Treat it that way. You would not give a brand-new joiner unsupervised production access on day one. You would not let them act on customer accounts without identity, authorisation, audit, supervision, and a clear performance review. Agents need the same. Most of the operational burden in this book follows from that single framing.
This is a book for environments where failure is not an option. Heavily regulated industries (financial services, insurance, healthcare, life sciences, government, critical infrastructure, defence). Industries where reputational damage cuts deep (consumer brands operating at scale, professional services, anything with safeguarding obligations). Industries where a single incident can produce regulatory fines, customer compensation, mandatory disclosure, or licence loss, particularly anything involving leaked data, exposed PII, or a system compromise that makes the news. The patterns and controls in this book are calibrated for those settings.
If your context is different, you should still read it. Then make a judgement call. Most of the controls described here are layered: identity, policy engines, asymmetric guardrails, immutable audit, AAS-style red teaming, two-person rule for high-tier tools, formal Agent Contracts, multi-stage approval paths. In a non-regulated environment with a small team and limited blast radius, applying every layer is over-engineering. The right move is to take the architecture (the seven layers, the trust boundaries, the build-pillar plus runtime-pillar plus closed-loop synthesis in Chapter 20) and dial the operational rigour down to match your actual risk.
Three rules of thumb for scaling the controls down:
Keep the architecture; relax the artefacts. The seven-layer reference, the OBO identity pattern, the policy gate, the trace-per-turn observability, the eval set: keep all of them. They are not expensive to design in from the start; they are expensive to retrofit. The artefacts around them (the formal threat model, the multi-party approval record, the documented quarterly review, the dual-key rotation cadence) can be lighter for low-stakes deployments.
Tier the agent honestly, then size the controls to the tier. The four-tier scheme in Chapter 3 works in non-regulated contexts too. A Tier 1 internal copilot in a marketing team does not need a Tier 4 control set. A Tier 4 customer-facing agent needs a Tier 4 control set wherever it lives.
Re-check the calculation when blast radius changes. A non-regulated agent that grows new tools, new data scopes, or new user populations is a different agent. Re-tier and re-control accordingly.
If your environment is regulated, the operational burden in this book is approximately what regulators and auditors will expect of you, and reading it as a checklist of mandatory practice is reasonable. If it is not, read it as a menu and pick what your risk profile justifies.
What you will not find here
I want to be honest about scope, because the AI literature is full of books that promise everything and deliver an extended product brochure.
You will not find a vendor comparison. The frameworks change too fast for a printed book to be useful as a buyer’s guide. You will find architecture patterns and the questions to ask any vendor, but you will not find a leaderboard.
You will not find prompt engineering basics. There are good resources for that. This book starts where prompt engineering ends, at the point where prompts become part of a system that does things.
You will not find a survey of every agent framework. I will name a few when they make a point clearer, but the book is not built around any single one. The patterns matter. The framework underneath them is replaceable.
You will not find AGI speculation. I have nothing useful to add to that conversation. The systems we build now are powerful and limited and need governance. That is enough material for one book.
You will not find a bullet point list of “challenges and opportunities” at the end of any chapter. If a chapter has challenges, they are described in the body where they belong.
How to read this book
The six parts are designed to build on each other, but they are also reasonably self-contained.
Part I reframes the risk model. If you have not yet had the argument with a stakeholder about why agents are different from chatbots, read it first. It gives you the vocabulary.
Part II is the architecture core. If you are designing a new agent today, this is where the load-bearing material is. Read all four chapters before you write a design doc.
Part III is for teams that have an agent in pilot and are getting nervous. Evals and observability are what turn a hopeful pilot into a system you can defend.
Part IV is for anyone whose agent will read enterprise documents. Most agents will. Skipping this part is how teams end up in the news.
Part V is for engineering leaders dealing with coding agents. The chapters apply whether you call your tool Copilot, Cursor, Claude Code, Codex, or something internal. The principles are the same.
Part VI is the governance and rollout layer. Read this if you are the person who will sign the production change ticket. Read it twice if you are the person who will explain the incident.
Each chapter follows the same shape. A short opening that sets up the problem in plain language. Learning objectives split into “will understand” and “will be able to”. The body of the chapter, broken into named sections. A worked example, which in this book is almost always drawn from financial services because that is what I know best. Two or three exercises at the end, designed to be done in an afternoon by a small team. A summary that tells you what to remember if you forget the rest. And footnoted notes that point at primary sources.
A word on examples. The book uses a mix of real-public-record cases, anonymised case studies from work I have been close to, and fully synthetic scenarios. I will mark which is which. Where I am drawing on a project I cannot name, I will say so and stick to the parts I am allowed to discuss.
Downloading the example code files
The code bundle for the book is hosted on GitHub at https://github.com/PacktPublishing/Enterprise-AI-Agent-Security. We encourage you to download or clone the repository and use the code alongside the book.
A note on tools and versions
The book is written in 2026 and reflects the state of agentic AI tooling at that time. MCP is the canonical tool protocol I assume. Claude, GPT-class models, and a small number of open-weights models are the assumed substrate. Where I name a specific product, I will pin it to the version. Where I describe a pattern, I will try to make it survive the next two model generations.
If a specific framework I name has changed by the time you...




