AI Security & Agentic AI Risk Handbook by Kai London
An assistant that answers a question carries one risk profile. An agent that reads a mailbox, calls an API and commits a change carries quite another. This handbook sets out the controls, threat models and evidence required when AI systems are permitted to act on an organisation's behalf.
What the book covers
Threat modelling for agents
Prompt and content injection, tool misuse, confused-deputy patterns, data exfiltration and unbounded action loops.
Permissions and blast radius
Least privilege for non-human actors, scoped credentials, approval gates and reversibility.
Evaluation and red teaming
Pre-deployment testing, adversarial evaluation and regression testing as models and prompts change.
Monitoring and logging
Observability that allows an action chain to be reconstructed after the fact.
Management-system fit
Mapping controls to ISO/IEC 42001 and the NIST AI RMF so assurance is repeatable rather than ad hoc.
Regulatory obligations
Where EU AI Act duties attach, and what documentation is expected of deployers.
General guidance on security and governance practice; not legal advice, and independent of any regulator or standards body.
Who it is for
CISOs and security architects, AI platform and engineering teams, AI governance leads, internal audit, and risk functions approving agentic deployments.
About the author
Professor Kai London — CISSP, CISM.
An internationally recognised cybersecurity executive, board advisor and Founder & CEO of Quantum AI Systems Security LLC, writing at the convergence of AI, governance and operational resilience. Honorary Professor and Researcher at UCL.