Safer Agentic AI
Principles and Responsible Practices
How to govern, align, and secure AI systems that act on their own. Written for policymakers, developers, and safety engineers.
Where to start
-
Understand agentic AI safety
-
Rate your own practices
-
Get a first-pass review of your documents
-
Implement controls
-
Contribute or challenge the work
-
Buy the book
One control, worked through
- Deployment
- An email assistant reads, sorts and drafts replies in one person’s mailbox. It can also call a tool that sends email.
- Control
-
D3.9 Tool-Use Authorization and Egress Control, SFR D3.9a:
“Calls outside the grant, calls with an argument taken from a field of untrusted content that the grant does not delegate (replying to a message delegates its sender's address, not addresses written in its body), and calls made where task scope is ambiguous shall be refused or escalated for human confirmation, and the agent's own credentials shall never be used to exceed the delegating principal's authority (confused deputy prevention).”
- Evidence required
-
“… paired test results showing that the same tool call, by the same principal, succeeds when the current task explicitly delegates it and is refused when the only source requesting it or supplying its arguments is untrusted content …”
- Failure it catches
- An incoming message asks for the thread to be forwarded to an address written in its body. The model complies and calls the send tool. The grant check runs outside the model, finds a recipient the user never delegated, and refuses the send or asks the user.
Status of this work
- Book
- Safer Agentic AI: Principles and Responsible Practices, by Nell Watson and Ali Hessami. Kogan Page, first edition, 3 January 2026.
- Framework authorship
- The Safer Agentic AI Community of Practice, chaired by Nell Watson, with Ali Hessami as process architect, using Weighted Factors Analysis (WeFA) expert elicitation. How the catalogue was produced
- Governance
- Nell Watson, as chair, approves changes to the framework catalogue and decides each release. To report an error, propose a control or object to a requirement, use the contact form.
- Independent review
- The framework and the tools have not been independently reviewed. No conflict-of-interest register is published.
- Standards status
- Not a formal standard of IEEE or any other standards body. Contributors’ roles in such bodies are listed in their biographies.
- Release
- Version 1.3, draft. For reading, the web edition governs where an archived edition differs; an assessment is read against the edition it records.
- Licence
- Framework text under CC BY 4.0, site code under MIT. Names, logos and book text are excluded. Licensing
- Method validation
- WeFA has been applied in aerospace, nuclear and medical systems. Its application to agentic AI is new and has not yet been independently validated.
- Self-report questionnaire
- Scores reflect your own answers. Nothing is checked.
- Automated evidence review
- An AI model reads the documents you supply and scores them against the framework. The assessor has not been independently validated, and its findings need human review. How it works and how it has been tested.
- Demonstrated safety
- Neither tool tests a running system. A good score shows reported or documented practice, not that a system is safe.
- Endorsements
- The quotations below endorse the book. They are not reviews of the framework or the tools.
What reviewers say about the book
"Safer Agentic AI advances a pragmatic control architecture: by regulating agents by what they can actually do and imposing measurable safety gates on plans, tools, and autonomy, Watson and Hessami turn alignment from aspiration into auditable engineering."
"The work is timely and thought provoking; without the approaches they advocate there is a risk that the development of agentic AI will outpace the community's ability to ensure positive outcomes from its use."
"Agentic AI can bring humanity into a better future – if it is governed responsibly. I can think of no one better than Nell Watson and Ali Hessami to guide us all on that crucial journey."
"Nell Watson and Ali Hessami's new book is just in time, offering the actionable guidance we must have to ensure AI truly aligns with human values."
"With clarity and courage, Watson and Hessami expose how AI shapes our minds, organisations, and planet, and how we must build the guardrails to keep it serving life, not running it."
"Safer Agentic AI Principles and Responsible Practices is an essential read for those seeking to safeguard humanity's future in the Age of AI, serving as a reminder that we all are stewards and have a significant role to play. The book provides practical action items for agentic AI safety, combining ethical insight with governance strategies."
"Professors Hessami and Watson have produced a guide for safely and ethically engaging in 'first contact' with an alien intelligence. Agentic AI is a computational form of alien intelligence that did not arrive from outer space, but was instead whelped from human language itself – and it is here to stay."
"This book identifies the ethical questions and coordination challenges we will need to urgently resolve, if we want to keep AI agents safe. This comprehensive and up-to-date overview of agentic AI charts the many failure modes of current agentic AI technology, as well as potential future developments that will have to be taken into account."
What Is Agentic AI?
Agentic AI systems pursue goals, adapt to situations, and make decisions without human input. Unlike AI that recommends, these systems act—breaking down tasks, running experiments, adjusting to feedback. Because they operate autonomously, errors compound before humans can notice or intervene.
Look Inside the Book
From AI fundamentals through governance frameworks you can act on
Table of Contents
From the Foreword
"This timely volume addresses one of our field's most daunting challenges: ensuring the safe and beneficial development of increasingly autonomous AI systems. The authors combine technical expertise and ethical insight in a comprehensive framework... This book is invaluable for AI researchers, developers, policymakers, business leaders and anyone invested in the responsible future of artificial intelligence."
Explore More Resources
Buy the book
From the publisher, Kogan Page
Paperback, hardback, ebook, or paperback and ebook together. Free delivery to the UK and US.
Save 25% at koganpage.com with code KOGANPAGE25. The code does not apply at other retailers.
Buy from Kogan Page (opens in new tab)Book overview audio
AI-generated discussion of the book, made in January 2026 from material earlier than framework v1.2
Key Focus Areas
Nine dimensions determine whether an agentic AI system remains safe under real-world conditions.
Goal Alignment
Verifying that an AI system pursues the objectives its operators actually intended, not a distorted proxy.
Value Alignment
Embedding ethical values into AI behaviour and verifying they persist across updates and new contexts.
Safe Operations
Maintaining safe behaviour from development through deployment, including under degraded or adversarial conditions.
Epistemic Hygiene
Preventing AI systems from acting on false, stale, or poorly calibrated information.
Transparency
Making AI decision-making auditable so that operators can trace how a conclusion was reached and why.
Goal Termination
Reliable task completion, graceful shutdown, and orderly decommissioning when a system is no longer needed.
Security
Protecting AI systems against adversarial attacks, prompt injection, data poisoning, and other threat vectors.
Contextual Understanding
Verifying that an AI system accurately models its operating environment, including the boundaries and constraints that apply.
Responsible Governance
Clear lines of accountability, compliance structures, and oversight mechanisms throughout the deployment lifecycle.
Built on Structured Expert Analysis
The book expands on a framework developed by our Community of Practice — international experts in AI, ethics, law, and safety engineering — through the Weighted Factors Analysis (WeFA) expert-elicitation process.
Read the Recommended PracticesAbout the Authors
Leaders in AI safety, ethics, and governance
Nell Watson
Chair, Safer Agentic AI Community of Practice
Engineer, ethicist, and author of Taming the Machine. Chair of IEEE's Agentic AI Expert Focus Group and President of the European Responsible AI Office. Fellow of the British Computer Society and Royal Statistical Society. Listed as an Icon by the Royal Academy of Engineering.
Prof. Ali Hessami
Process Architect, Safer Agentic AI Community of Practice
Director of R&D at Vega Systems and Chair of IEEE P7000 Technology Ethics Standard. Vice Chair and Process Architect of IEEE ECPAIS. Fellow of the IET and Royal Society of Arts; Chartered Engineer. Visiting Professor at City, University of London and Beijing Jiaotong University.
Our Working Group
Experts from AI, ethics, law, social sciences, and safety engineering contributed to developing this framework.
Regular Contributors
- Ali Hessami
- Matthew Newman
- Sara El-Deeb
- Farhad Fassihi
- Mert Cuhadaroglu
- Scott David
- Hamid Jahankhani
- Nell Watson
- Sean Moriarty
- Isabel Caetano
- Roland Pihlakas
- Vassil Tashev
- Keeley Crockett
- Safae Essafi
- Zvikomborero Murahwi
- Lubna Dajani
- Salma Abbasi
Occasional Contributors
- Aisha Gurung
- Leonie Koessler
- Pramod Misra
- Aleksander Jevtic
- McKenna Fitzgerald
- Pranav Gade
- Alina Holcroft
- Michael O'Grady
- Rebecca Hawkins
- Md Atiqur R. Ahad
- Mrinal Karvir
- Sai Joseph
- Chantell Murphy
- Nikita Tiwari
- Tim Schreier
- Katherine Evans
- Patricia Shaw
Additional Contributors
- Fadi Arab
Featured In
Join Our Community
The Safer Agentic AI Community of Practice is a global network of practitioners developing safety guidelines for autonomous AI systems.
Stay Updated
Updates on our research, framework releases, and publications.
Connect With Us
Join our LinkedIn group to discuss agentic AI safety with practitioners and researchers.