Short answer: Securing AI agents means controlling what they do, not only what they say. This playbook covers the full program in 10 chapters: inventory and identity, least-privilege access, runtime policy on every tool and API call, prompt-injection defense, audit evidence, and mapping to the OWASP Agentic Top 10 2026, NIST and the EU AI Act.
How to check every agent action against policy before it runs, and what a runtime control set must include.
Why per-app guardrails leave gaps, and how to govern a whole agent fleet with one inventory and one policy layer.
All ten ASI risks in plain terms, how they relate to the 2026 LLM Top 10, and the four controls to start with.
The four layers every agent needs before launch: input checks, tool scope, runtime policy and human approval.
Why filters fail against indirect injection, and how trust boundaries and action controls contain it (ASI01, LLM01).
Per-agent identity, default-deny tool access, time-bound permissions and a runtime check on every call (ASI03).
What to log for every action, how to trace multi-agent chains, and the evidence auditors ask for.
One control map across four frameworks, with the EU AI Act dates as amended in 2026.
Trust boundaries, handoff authorization and cascade containment for agents that delegate (ASI07, ASI08).
Four detection signals, a registry checklist and a response plan for unapproved or drifting agents (ASI10).
Where to start: new to agent security → chapters 1, 3, 4. Building controls → 6, 5, 9. Proving them → 7, 8. Finding what you missed → 10.
Controlling what autonomous agents can access and do: their identity, the tools and data they reach, and each action they take, with a record of every decision. It extends application security from model output to real-world actions.
Build an inventory of every agent and its owner, give each agent its own identity with default-deny tool access, then add runtime policy on high-risk actions such as payments, deletes and data exports.
The OWASP Top 10 for Agentic Applications 2026, the OWASP Top 10 for LLM Applications 2026, NIST AI RMF and NIST's AI Agent Standards Initiative, and the EU AI Act as amended by the 2026 AI Omnibus.
No. Prompt instructions guide behavior but cannot stop an action. Enforcement has to sit outside the agent and check each tool or API call before it executes.